# Genuine STAR for R3 / Bar Raiser Notes for Adarsh Vishwakarma. Job **10454435** still names **none**. R1 (13 Aug 2026) and R2 (18 Aug 2026) are done. This fragment is spoken STAR for a **third live** and a possible **fourth** (HM / Bar Raiser). Other SDE I / UTA / AUTA candidates reported these prompts. Your loop may differ. Unnamed stays unnamed. Bar Raisers bust fabricated stories. The projects below are genuine. The numbers are only those on the Aug 2026 resume. If they already heard Horizon 17th or Ylogx 99.9% in an intro, start on the **defect** inside that story, or switch primary per the spread — do not retell the win. **Java once, then stop.** Live Code is Java. Production was Python (FastAPI, LangGraph, YOLOv9, ROS2), TypeScript (NestJS, React), and ROS2 C++/Python on Horizon. Do not pretend Ylogx or IQVIA ran on the JVM. **InstaRecon / PhiSiFi.** Not an LP story. One line: security-awareness demo, consent, no production attacks. Then StratifyLabs, Argus, Ylogx, or IQVIA. --- ## 0 How to speak Situation 15s, Task 10s, Action 40s with three mechanisms, Result 15s, STOP. They will 10x, ask what YOU did, ask what you’d do differently, ask how you measured. If you don’t remember a number: “I don’t remember the exact protocol; the resume number is X; the mechanism was Y.” Never invent a baseline. Use I. No we-washed credit. Use **we** only for a team result you could not have produced alone (Horizon’s 17th place is a team score). Ylogx RLS in Postgres is an I-designed-and-shipped claim. **Clock.** A Bar Raiser will interrupt. Four paragraphs, then shut up. Unnamed LP blocks on other AUTA loops were often 10–20 minutes with two or three prompts; some BRs ran ~30 minutes on behavioral. If you run in four-word bullets they think you memorized a poster. If you run past two minutes they cut you. **Why-chain.** They will ask “why?” about three times on the **same** story. Answer the why that was asked. Do not restart the STAR. Do not change project unless they change Leadership Principle. 1. Why that approach instead of the obvious one — name the discarded option and the failure mode. 2. How did you know it was working — a resume metric or an eval, not a feeling. 3. What would you do differently — a real missing test or a sequencing miss, not false humility. **Front-load the defect.** If you open with “we placed 17th,” they have nowhere to probe except “nice.” If you open with “the costmap was planning on stale occupancy while the operator feed still looked live,” they can stay for five minutes. That is the round. **Do not invent.** Missed ERC. Missed 99.9%. Manager fight. Mentee count. AWS bill. Qdrant. Haar cascades. p95. Cache TTL. Student headcount. SEO 403 as Dive Deep lead. Fabricated intern drama. If it is not on the Aug 2026 resume, it is not in this room. --- ## 1 Genuine stories — one primary per LP (16) Primary spread is locked. Ylogx already has three primaries (Ownership, Frugality, Dive Deep). IQVIA three (Invent, Learn, Backbone). Argus three (Are Right, Highest Standards, Success and Scale). Horizon three (Bias, Deliver, Earth’s Best). GiftedBooks two (Customer, Earn Trust). StratifyLabs is Think Big primary. IEDC is Hire and Develop primary. Do not promote a backup into a fourth primary. | LP | Primary | Backup (one line) | | --- | --- | --- | | Customer Obsession | GiftedBooks | Ylogx SQL RAG +65% analysis productivity | | Ownership | Ylogx RLS + 99.9% | Horizon GStreamer + costmap as ship-blockers | | Invent and Simplify | IQVIA Hybrid RAG + test-case gen | StratifyLabs browser inference −30% | | Are Right, A Lot | Argus 73%→89% mAP | Ylogx measured bot DB latency −35% | | Learn and Be Curious | IQVIA LangGraph | Horizon ROS2 / GStreamer / ZED | | Hire and Develop the Best | IEDC CUSAT Tech Team | Horizon teammates on launch order | | Insist on the Highest Standards | Argus 15k + mAP gate | Ylogx 99.9% / sub-210 ms | | Think Big | StratifyLabs marketplace + 3D lab | IQVIA 200+ ranked sites | | Bias for Action | Horizon ERC date | Eight hackathons / Redis ship | | Frugality | Ylogx Redis −35% + CloudFront/ECS | Horizon costmap vs extra hardware | | Earn Trust | GiftedBooks RAG + 99.5% | Ylogx three-tier RLS (JOIN leak) | | Dive Deep | Ylogx Redis + RLS | IQVIA hybrid retrieval + LangSmith | | Have Backbone; Disagree and Commit | IQVIA hybrid vs vector-only | Ylogx RLS-in-DB vs app-only | | Deliver Results | Horizon 17th/80+ | Ylogx 40% faster reports / 99.9% | | Strive to be Earth’s Best Employer | Horizon knowledge sharing | IEDC CUSAT Tech Team | | Success and Scale Bring Broad Responsibility | Argus 20+ cameras after 89% | Ylogx ALB/ECS with RLS already on | If they do not name the LP, lead Dive Deep then Deliver Results. Third: GiftedBooks 3–10 minutes. If intro already used Horizon, switch Deliver Results to Ylogx 40%/99.9%. Aditya-style trio (three prompts, three projects): Ownership = Ylogx, pressure = Horizon, learn = IQVIA. Do not reuse Ylogx three times. --- ### Customer Obsession — primary GiftedBooks **Primary project:** GiftedBooks VR learning suite — RAG over the student’s PDF, PYQ topic ranker, hours → 3–10 min. **Backup project:** Ylogx SQL RAG chatbot: non-technical users, +65% data-analysis productivity, behind three-tier RLS. **Spoken STAR** Situation. GiftedBooks is a VR learning suite: 3D labs, AI avatars, and a study path. The customer is a student who already has PDFs and still loses hours on doubts. A pretty lab that cannot answer a question from *their* material is decoration. A generic chatbot that sounds like a textbook is also the wrong customer. They are trying to pass an exam with their notes, not with a model’s prior. Task. I needed to cut time-to-answer without inventing content, and to make the assistant usable during study — not only in a demo. That meant grounded Q&A plus a study path that was honest about what exams actually ask. Action. I shipped RAG over uploaded PDFs so retrieve-then-generate ran on *their* chunks, not a stuffed context window. I kept generate small so the API stayed under 300 milliseconds; chunks too large blow that budget, chunks too small miss the table on page forty. I held 99.5 percent uptime as a product constraint, not vanity. I added a PYQ topic ranker as analytics *separate from the LLM*, so people studied what exams ask instead of an LLM guessing the syllabus. Per-user PDFs must not retrieve another student’s notes. Embed once per PDF; do not re-embed on every question. Result. Average doubt resolution went from hours to three to ten minutes. Reading efficiency improved 35 percent. Engagement improved 50 percent. Content comprehension moved 2.5 times. Sub-300 millisecond API. 99.5 percent uptime. GitHub currently shows an AegisAI README on that repo; I will not mix that product into this story. **Why-chain** 1. Why RAG, not dump the whole PDF into the prompt? Tokens, invented citations, and you miss the sub-300 millisecond budget. The student memorizes a fluent lie. 2. Why PYQ as a separate analytics job, not an LLM syllabus guess? Honesty about what exams ask. That is working backwards from the customer’s actual exam, not from a chatbot voice. 3. Why 99.5 percent and sub-300 milliseconds as product, not ops trivia? A study assistant that times out during revision is not a product. The customer is in a session, not a demo. **BR follow-ups** 1. **What you did vs team.** I owned the RAG path and the PYQ ranker. VR labs and avatars were already the surface. I will not claim I designed every 3D lab. The failure I named out loud was “decoration without answers.” 2. **Metric how-measured honesty.** Hours → 3–10 min is the resume number. I do not have a student headcount. I will not invent MAU, a gold-set size, or LangSmith on this product. If they ask how I measured 35/50/2.5×: those are the resume numbers; I will not invent a survey protocol I do not remember. 3. **What you’d do differently.** I would add a small wrong-citation versus grounded eval — the IQVIA habit. That eval is **not** on the GiftedBooks resume bullet. I would add it. I would not pretend I already shipped it. 4. **10×.** Embed once; cache retrieval; keep generate small; per-user isolation so you do not retrieve another student’s notes. Do not invent Kafka. Do not steal LangSmith onto this bullet. 5. **Conflict.** Product backbone, not a fight: we will not answer off-PDF. If the student asks a definition that is not in their upload, the assistant abstains. That is pushback on a customer request that would have been a hallucination. See §3. 6. **Failure inside the same story.** First pass without PYQ still left people guessing *what* to study. A pretty lab with slow or generic Q&A failed first. Pretending the VR surface was enough was the self-critical beat (Earn Trust uses the same product; do not retell this as a fifth GiftedBooks loop — GiftedBooks is already at two primaries). **What you will NOT say.** Student headcount. SEO. AegisAI features from the wrong README. LangSmith on GiftedBooks as if it shipped. Zero-incident claim. “We never hallucinated.” A named outage. Qdrant. --- ### Ownership — primary Ylogx RLS + 99.9% **Primary project:** Ylogx intern (Nov 2024–Oct 2025) — three-tier RLS + RBAC in Postgres on the path the bot and reports share; 99.9% held. **Backup project:** Horizon: GStreamer 60 FPS and the costmap as ship-blockers; 17th of more than eighty. **Spoken STAR** Situation. At Ylogx I interned on a full-stack AI BI product: FastAPI and NestJS over Postgres, React dashboards, and a LangChain SQL RAG chatbot that turned natural language into SQL. The chatbot and the report builder would have been wrong *and* a leak if organizational boundaries were only an application check. On a demo, generated SQL returned rows. That looked like ownership of the ticket. It was not ownership of the product. Task. I was not the named security team. I was the intern shipping the data path. I owned correctness and isolation for three organizational tiers, not “chatbot works on my user.” I never treated “filter in the UI and ship” as done. Action. I put row-level security and RBAC for three organizational tiers in Postgres so generated SQL ran as the user’s database role, not as a service superuser. A JOIN could not outrun the app because the database refused the row. NestJS owned auth and report CRUD; FastAPI owned NL → constrained SQL; the LLM never held a superuser connection string. I put Redis on the hot bot path for a 35 percent drop in database latency so security did not mean “add a check and ignore load.” Cached answers still had to be the right tier. I kept serving on CloudFront, ECS, Docker, ALB, Route 53 / GoDaddy DNS so the BI path that already had RLS was the one that stayed up. Result. 99.9 percent uptime held while report generation went 40 percent faster. Analysis productivity for people who were not writing SQL went up 65 percent. Bot database latency dropped 35 percent. Thirty KPI dashboards, plus 60 percent ops efficiency, sub-210 milliseconds on the dashboard path — that path is not an LLM. **Why-chain** 1. Why RLS in the database, not app-only `WHERE org_id = ?`? A JOIN outruns a UI filter. Defense in depth. Generated SQL that is legal for the intern account is still a leak in production. 2. Why still Redis after RLS? Isolation is not an excuse for load. Repeat NL questions and schema lookups are cache-shaped. Minus 35 percent is the resume number. 3. Why is this Ownership and not just a security ticket? Because “that’s not my job” would have been “the chatbot demoed.” Owners act on the path the bot *and* the reports share. Leave-things-better is Success and Scale, not this LP. **BR follow-ups** 1. **What you did vs team.** I designed and shipped RLS policies + role mapping on the generated-SQL path, and Redis on that same path. I was an intern, not a manager, not cluster owner. I will not claim I ran the AWS account. 2. **Metric how-measured honesty.** 99.9 percent, 40 percent, 65 percent, minus 35 percent, 30 dashboards, 60 percent, sub-210 ms are resume numbers. I will not invent p95, a cache TTL, or an AWS bill. If I don’t remember the exact latency protocol: the resume number is minus 35 percent bot database latency; the mechanism was Redis on repeat NL and schema lookups, still behind RLS. 3. **What you’d do differently.** Treat RLS policies as tested artifacts from day one — denied cross-tier cases — the same way I later treated Argus mAP as a gate. Write cache invalidation rules down with RLS. The resume does not detail TTL; I will not invent one. 4. **10×.** More cache hits on repeat questions (still tier-scoped); connection pooling; CloudFront for static; RLS must still hold or you scale leaks. Maybe read replicas for dashboards later. I did not shard. I will not invent Kafka or QPS. Production path I will defend is ECS + Docker + CloudFront. 5. **Conflict.** Technical: RLS-in-DB versus app-only filters. I would not ship the bot on app-only checks because a JOIN leaks. No fight scene. Commit was three-tier RLS and RBAC as a platform rule, not a UI hide. 6. **Failure inside the same story.** The chatbot “worked” while every natural-language turn hit Postgres and a JOIN could leak. Anecdote said done. Metrics said latency and isolation were the product. App-only role checks fail closed in a demo and open in a JOIN. **What you will NOT say.** Missed 99.9%. AWS bill. p95. TTL. SEO 403 as this story. Qdrant. “I was the security team.” Manager fight. Kubernetes-as-owner (k8s is on the skills list; I was not cluster owner). Login Tracker. Rate Limiter LLD. --- ### Invent and Simplify — primary IQVIA Hybrid RAG + test-case gen **Primary project:** IQVIA intern (Apr 2026–present, Kochi) — Hybrid RAG on 200+ page BRDs; LangGraph adaptive retrieval; LangSmith evals and test-case generation. **Backup project:** StratifyLabs browser inference; ML iteration −30 percent; marketplace 50+. **Spoken STAR** Situation. I am a software developer intern at IQVIA in Kochi. Analysts faced more than two hundred websites on one side and more than two-hundred-page BRDs on the other. Manual synthesis does not scale. Hand-written test cases from a two-hundred-page requirements document also do not scale. Stuffing the PDF into a context window blows tokens and invents citations. The fast “invention” is a vector index and a prompt pile. That is not coverage. Task. I had to make research and retrieval checkable: rank sources, retrieve the right clause, generate test cases a human can audit. Invention without a stop condition is another fluent paragraph. Simplification without a mechanism is a deleted feature. Action. I architected a LangGraph multi-agent Deep Research system: Firecrawl, Bing, DuckDuckGo, and Google Playwright in parallel, then a ranker across more than two hundred sites. State lives on the graph so a failed scrape retries instead of restarting the run — that is the checkpoint. I did not insist on one in-house scraper. I built Hybrid RAG on Azure AI Search with hybrid and semantic retrieval, plus a GraphDB, plus LangGraph choosing retrieval mode. Lexical search catches clause IDs like “R-141.” Semantic search catches “what is the retention policy.” Graph hops catch “what depends on this requirement.” Chunk by structure: clause IDs, tables, headings. The writer is citation-gated: no overlapping span, no sentence. Test-case generation is itself a traced LangSmith job. I discarded Qdrant; the resume is Azure AI Search and GraphDB. I discarded a linear chain because retries and parallel tools are a graph. Result. More than two hundred websites ranked. More than two-hundred-page BRDs processed with traces and generated test cases. No extra percentage on the resume. I will not invent one. **Why-chain** 1. Why hybrid plus graph, not one vector index? Two-hundred-page BRDs have clause IDs and requirement graphs that cosine misses. Similar embeddings are not coverage. Test cases from the wrong section waste whoever consumes the BRD. 2. Why four tool families, not one scraper? Each family fails on different sites. Parallel plus LangGraph checkpoint is cheaper than restarting the run. 3. Why LangSmith as the simplify, not “fewer prompts”? If a generated test case has no overlapping span with the retrieved section, the path is not done. That is the stop. Fluency is not a test. **BR follow-ups** 1. **What you did vs team.** I owned the graph of nodes, the retrieval mode choice, and the eval/test-case path. I will not claim I built Azure AI Search as a product. Production is Python FastAPI. 2. **Metric how-measured honesty.** 200+ sites, 200+ page BRDs, evals, test-case generation. No extra percent. No invented latency SLA. No gold-set size I will invent. LangSmith traces are how I know a path ran; overlapping span is how I know a sentence is allowed. 3. **What you’d do differently.** Freeze an eval slice of BRD sections to expected test cases before adding a fourth search tool. Citation-required writer in the type system earlier: no span, no sentence. Freeze gold questions with expected citations before adding Playwright. 4. **10×.** Do not re-embed a BRD on every question; hybrid-retrieve a small set of chunks, not 200 pages into the prompt; graph hop is cheaper than another LLM call when the question is relational; cache scrapes; fewer Playwright renders, more Bing snippets when cost spikes. Secrets in env, not in traces; do not log full BRD text in LangSmith in production. I will not invent a multi-tenant model the resume does not state. 5. **Conflict.** Hybrid versus vector-only. A teammate who wants vector-only is optimizing for a ship date. I stated similar embeddings are not coverage. Once we chose the path, I committed on evals. Disagree on retrieval. Commit on LangSmith. No named interpersonal incident. 6. **Failure inside the same story.** Vector-only retrieves similar paragraphs and misses “R-141.” First-hit ranking looks confident when two high-ranked pages disagree. Empty scrape still feeding the writer is the LangSmith-visible bug. Tutorial-complete without an eval failed first. **What you will NOT say.** Qdrant. Extra percent. p95. A foundation model I trained. Kafka. JWT/RLS theatre (that is Ylogx). “I invented LangGraph.” Manager quote. --- ### Are Right, A Lot — primary Argus 73%→89% **Primary project:** Argus industrial computer vision — YOLOv9, 15k images, 73→89% mAP, 24 FPS, then 20+ cameras. **Backup project:** Ylogx measured latency: 40 percent faster reports, sub-210 ms, Redis −35 percent — instinct said the model was slow; the metric said the database was. **Spoken STAR** Situation. Argus is industrial computer vision for PPE compliance and attendance on camera. A detector that looks fine in a notebook is not right on a floor. At 73 percent mAP you miss helmets and you also fire junk boxes. Either failure trains people to ignore the alarm. Instinct said the training curve was going the right way. The metric said it was not a number I would page a supervisor on. Task. I had to disconfirm the notebook. “Right” is an eval you would alert a human on, not a screenshot. I treated 73 percent mAP as a fail, not a starting blog metric. Action. I trained and evaluated on more than 15,000 images and iterated YOLOv9 until 89 percent mAP. Non-maximum suppression sits on the boxes so junk duplicates do not become extra alarms. I required 24 FPS so the model we called “right” could keep up with the stream — a heavier two-stage detector that dies at twenty cameras is not right either. OpenCV for decode and resize, not as the detector. I wrote events to Postgres so compliance was an audit trail, not a GIF. Video stays off the database. I did not replace boxes with an LLM looking at frames. I did not scale 73 percent to twenty cameras to look busy. Result. Seventy-three to 89 percent mAP, 24 FPS, more than 15,000 images. Then more than twenty camera feeds. Safety violations down 50 percent. Compliance doubled. After the gate, not before it. **Why-chain** 1. Why not ship 73 percent? Misses and junk boxes train people to ignore alarms. That is being wrong at scale, not “early.” 2. Why YOLO plus NMS at 24 FPS, not a heavier detector or an LLM on frames? Twenty-plus cameras is the constraint. Fluency is not mAP. Boxes must be auditable. 3. Why Postgres events, not video files as the source of truth? Audit trail of what fired. Events, not blobs. How I knew I was right later: 89 percent mAP, 24 FPS, then −50 percent violations and 2× compliance. **BR follow-ups** 1. **What you did vs team.** I owned training, the mAP/FPS gate, NMS, and event logging. I will not invent a plant-floor headcount or a named safety manager quote. GitHub `argus-stream-api-server` README is 404 — no invented routes. 2. **Metric how-measured honesty.** mAP is the eval I will defend. 73 → 89, 15k+ images, 24 FPS, 20+ cameras, −50% violations, 2× compliance. I will not invent precision/recall beyond mAP, a train/test split the resume does not name, or 95 mAP. If they ask how 50 percent violations was measured: the resume number is −50 percent after the gate; the mechanism was alerts from the 89 percent / 24 FPS model plus Postgres logs. I do not remember a named protocol beyond that; I will not invent one. 3. **What you’d do differently.** Hold out a camera earlier so 89 percent is not an accident of leak. The resume does not name that split; I would add it; I am not claiming I already published it. Per-camera error budget is would-add, not claimed. 4. **10× cameras.** Do not scale 73 percent mAP. Workers per stream or batch. Do not run twenty copies of the fattest model if 24 FPS breaks. mAP gate then multiply. Not EKS (Ylogx is the ECS story). Not Kafka. 5. **Conflict.** Scaling versus the gate. Someone who wants cameras live this week is not foolish. I still would not multiply a 73 percent model. That is Are Right overlapping Success and Scale — keep this card on the eval; scale is the third Argus primary, do not mash them into one sentence unless they ask both. 6. **Failure inside the same story.** Treating 73 percent mAP as progress. That is also the mistake / critical-feedback card. A missed helmet in odd lighting is a silent fail. A false box is a loud fail. Both live inside 73 percent. A one-line color threshold does not take PPE from 73 to 89 at 24 FPS. Haar cascades are **not** on the resume; I will not say we tried Haar first. **What you will NOT say.** Haar. YOLOv8 + ByteTrack + OSNet (Purplle-style, prep-only). Manager quote. 95 mAP. Precision/recall I did not publish. EKS. Student/plant headcount. “The feedback was my manager said…” --- ### Learn and Be Curious — primary IQVIA LangGraph **Primary project:** IQVIA — LangGraph state, hybrid Azure retrieval, LangSmith pass/fail, including generated test cases. **Backup project:** Horizon ROS2 / GStreamer 60 FPS / ZED 2 at 2M+ pts/s; 17th of more than eighty. **Spoken STAR** Situation. Ylogx had been FastAPI, NestJS, Postgres, and SQL RAG. Calling an LLM from FastAPI was already a skill. IQVIA’s work was multi-agent research, hybrid retrieval, GraphDB hops, and tracing. LangGraph, LangSmith, and Azure AI Search hybrid were not the previous day job. Tutorial-complete is not architecture. Task. I had to learn enough to rank more than two hundred websites and to retrieve from two-hundred-page BRDs without inventing citations. Curiosity needs a stop condition or it is wandering. Action. I learned LangGraph stateful graphs so a failed scrape does not restart the whole run — planner, parallel tools, ranker, writer, checkpoint. I learned adaptive retrieval: lexical for clause IDs, semantic for policy language, graph hops for dependencies. I learned LangSmith as a stop condition, including generated test cases, so curiosity had a pass or fail. I discarded “RAG means one vector index.” I discarded a linear chain for Deep Research because retries and parallel tools are a graph. I drew the graph before I added a fourth tool. Result. The platform as on the resume: more than two hundred sites ranked, more than two-hundred-page BRDs, evals and test-case generation. Production is Python. Live Code is Java. **Why-chain** 1. Why a graph, not “I learned prompting”? Retries and parallel tools are a graph. A linear chain dies when one scraper fails. 2. Why hybrid, not more tutorials on embeddings? Long BRDs need lexical plus semantic plus graph. Clause IDs are not cosine. 3. Why traces as the stop, not a course certificate? If there is no overlapping span, the path is not done. That is how I knew learning had landed. **BR follow-ups** 1. **What you did vs team.** I learned by drawing the graph and instrumenting evals, not by collecting certificates. I will not claim I invented the research paper. 2. **Metric how-measured honesty.** 200+ and 200+ pages. No extra percent. No invented gold-set size. LangSmith fail-if-no-span is the check. 3. **What you’d do differently.** Freeze a gold set of questions with expected citations before adding a third search tool. Write a one-page “what we will not retrieve” list earlier so agents do not wander two hundred sites without a ranker. 4. **10×.** Same as Invent: cache scrapes, fewer renders, eval set has to scale or you multiply fluent garbage. Learning more tools without a freeze is not curiosity; it is drift. 5. **Conflict.** None interpersonal. The internal conflict is “add another Playwright render” versus “the eval is not green.” I committed to traces, not to tool count. 6. **Failure inside the same story.** Treating RAG as one vector index. Tutorial-complete without an eval. Empty scrape still feeding the writer until LangSmith showed it. **What you will NOT say.** Qdrant. A certificate as the result. Uncensored LLM / abliteration / LoRA / DPO / mergekit (off-resume, weak for Amazon). Haar. “I mastered LangGraph in a weekend” as a fake clock. --- ### Hire and Develop the Best — primary IEDC CUSAT Tech Team **Primary project:** IEDC CUSAT Tech Team while CSE at CUSAT, CGPA 8.42. Honesty: team member, not a people manager. **Backup project:** Horizon teammates — unblock camera and mapping so someone else can launch the GStreamer node. **Spoken STAR** Situation. I was on the Tech Team at IEDC CUSAT while doing CSE at CUSAT with a CGPA of 8.42 out of 10. New people joining club and hackathon work did not share one stack. If only one person could run the build, the next event failed when that person was on another laptop. Amazon’s Hire and Develop language is raise the bar, develop leaders, coaching. I will say the honest size: I was a team member, not a manager. I did not hire anyone. Task. Make the next person able to run the build and take the risky module, not collect a mentee count I do not have. Growing means they can ship without my laptop. Action. I shared environment setup, reviews, and “here is the failure we already hit” on tools we actually used. At eight national and regional hackathons I paired on the risky module instead of hoarding it. On Horizon I made ROS2, camera, and costmap knowledge usable by the rest of the rover team — launch order, GStreamer pipeline, ZED into the costmap — because a solo notebook does not place 17th. I did not run ratings. I did not run an Amazon-style hiring bar. Result. Honest outcomes: IEDC Tech Team on the resume; Horizon 17th of more than eighty as core software; first at CodeRecet, MLH Best Project, Magnathon 2.0 runner-up. CGPA 8.42 is education, not this LP’s metric. I have no mentee count. **Why-chain** 1. Why pairing on the risky module, not “I assigned tickets”? I had no reports. The bar I could actually raise was: someone else can launch the node. 2. Why Horizon in the same breath as IEDC? Backup mechanism: camera and mapping knowledge leaving one head. Earth’s Best Employer is the growth/empowerment angle on Horizon; this LP is the bar-raising / sharing angle on IEDC first. Do not collapse them into one sentence if they ask both — switch emphasis. 3. Why no headcount? Because I will not invent one. A Bar Raiser who asks “how many people did you develop?” gets: I do not have a mentee count; here is what I actually did. **BR follow-ups** 1. **What you did vs team.** Shared setup, reviews, pair on the risky module. I did not hire. I did not rate. I was not a team lead on paper. 2. **Metric how-measured honesty.** CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up, eight hackathons, Horizon 17th/80+. No junior headcount. No retention survey. CGPA is not this answer’s metric unless they ask education. 3. **What you’d do differently.** Write a ten-line runbook earlier. The resume does not claim a wiki; I will not say we had one. 4. **10×.** Not applicable as “10× reports.” Honest: more people can run the launch file. That is the scale of a student org. 5. **Conflict.** None invented. If two people wanted to hoard the camera node, I still paired. No fight scene. 6. **Failure inside the same story.** Siloed knowledge. One laptop had the working environment. The next event would have failed on a different machine. **What you will NOT say.** Mentee of N. “I hired.” Ratings. Amazon bar-raiser role. Fake people-management. Earth’s Best Employer mashed into the same paragraph as this one if they already used Horizon for Earth’s Best — keep IEDC as the primary here. --- ### Insist on the Highest Standards — primary Argus mAP gate **Primary project:** Argus — 15k images, 73→89% mAP, 24 FPS gate before 20+ cameras. **Backup project:** Ylogx 99.9 percent uptime, sub-210 milliseconds, CI/CD on ECS and Docker. **Spoken STAR** Situation. Argus sits on industrial safety. 73 percent mAP is a model card. It is not a standard you page a supervisor on. Alarm fatigue is the cost of a low bar. A model that is accurate in a notebook and slow on the stream is also a low bar. Defects do not get sent down the line. Task. Put a gate in front of scale. Problems stay fixed: do not “fix” 73 percent by adding cameras. Action. I used more than 15,000 images, not a toy set. I moved YOLOv9 from 73 to 89 percent mAP with NMS on the boxes. I required 24 FPS on the stream. I logged to Postgres and containerized a path that could support more than twenty cameras. I did not scale the 73 percent model to those cameras to look busy. Lighting and vest color are dataset work, not a one-line threshold. Result. Safety violations down 50 percent. Compliance doubled. The standard was the gate, not a paragraph in a README. 89 percent mAP, 24 FPS, 20+ cameras, after the gate. **Why-chain** 1. Why is 73 percent a standard failure, not “early progress”? You miss helmets and you fire junk. Either one trains people to ignore PPE alarms. That is a defect sent down the line. 2. Why 24 FPS as a standard, not only mAP? A right box that arrives late is a missed intervention. Twenty cameras make latency a standard, not a nice-to-have. 3. Why Postgres logs as part of the standard? Compliance is an audit trail. A GIF is not a standard you can inspect later. **BR follow-ups** 1. **What you did vs team.** I owned the gate: mAP, FPS, data volume, then scale. I will not invent a QA team I did not run. 2. **Metric how-measured honesty.** Same as Are Right. I will not invent a second eval family. Highest Standards is the *gate*; Are Right is *disconfirming the notebook*. Same project, different sentence. Do not recite 73→89 twice as if they were different events if they already heard Are Right — start on “I still would not page a human at 73” and the 24 FPS bar. 3. **What you’d do differently.** Publish fail cases — missed helmet, false box, timeout — as the standard, not only the happy metric. Held-out camera earlier. 4. **10×.** Gate then multiply. Backup Ylogx: a chatbot that hits Postgres on every NL turn is also a low standard; 99.9 percent and sub-210 ms are the serving bar. Do not promote Ylogx to a fourth primary; this is backup only. 5. **Conflict.** Ship-now versus the gate. Technical. No manager fight. 6. **Failure inside the same story.** Treating 73 percent as a ship bar. That is the honest mistake. I will not invent a missed SLA to make Highest Standards sound more dramatic. **What you will NOT say.** Haar. Missed 99.9% as a “standards miss.” Manager quote. 95 mAP. --- ### Think Big — primary StratifyLabs marketplace + 3D lab **Primary project:** StratifyLabs CV SaaS — 3D sim lab, browser inference, 50+ marketplace models/datasets, URDF/WebGL, Gemini RAG bots, −30% ML iteration. **Backup project:** IQVIA Deep Research across more than two hundred ranked websites. **Spoken STAR** Situation. Computer-vision work gets stuck in a local loop: one GPU, one notebook, one model, one person. Iteration *is* the product. Thinking small is “train on my laptop and demo.” StratifyLabs is a CV SaaS with a 3D simulation lab at stratifylabs.design. It is not a one-model demo. The GitHub README is default Next.js; I will not invent features from it. The resume is the product. Task. Make the default path reusable: a CV intern who is not a ROS expert should still load a URDF in the browser and try a pretrained model without waiting on a lab machine. That is Think Big relative to “train on my laptop.” It is not a fake TAM. Action. I built a 3D simulation lab for real-time experimentation and prototyping. I put inference in the browser so ML iteration dropped 30 percent — WebGL/WASM on the request path for small models; big train jobs off that path. I shipped a marketplace of more than fifty pretrained models and datasets, community profiles, and a URDF editor with WebGL. Gemini RAG bots sit in the simulated environment as 3D characters, conditioned on the scene. Deep RLS stays on Ylogx; I will not steal it here. No LangSmith on this bullet. Result. Iteration time down 30 percent. More than fifty marketplace items. I will not add a go-to-market number. **Why-chain** 1. Why browser inference, not Colab-per-person? Iteration time. −30 percent is time, not a cluster I will invent. 2. Why a marketplace of 50+, not one hero model? Reuse versus a private notebook. 50+ is a catalog, not 50 GPUs per user. 3. Why URDF in WebGL, not native Gazebo-only? CV people who are not ROS still prototype. Horizon’s Gazebo is a different product; I will not merge the stacks. **BR follow-ups** 1. **What you did vs team.** Resume work: sim lab, browser inference, marketplace, URDF editor, Gemini RAG bots in-scene. I will not invent users from the default Next.js README. 2. **Metric how-measured honesty.** −30% ML iteration, 50+ models/datasets. No GTM. No user count. No cluster size. 3. **What you’d do differently.** Clearer “marketplace versus your fine-tune” boundary earlier. Eval harness for Gemini bots grounded in *current* sim state. Not claimed — I would add it. 4. **10×.** Catalog fetch, not a dedicated GPU per visitor; keep train jobs off the request path; auth on profiles so a RAG bot does not exfiltrate another user’s datasets. Do not invent Stratify RLS. Do not invent Kafka. 5. **Conflict.** Local GPU habit versus browser path. Technical. “Works on my GPU box” is the honest bug class. I will not invent a Jira ticket or a crash percent. 6. **Failure inside the same story.** Local-only inference made every experiment a machine problem. Browser memory / model size is the hard part I will name, not a fake outage. **What you will NOT say.** User count. TAM. AWS bill. Merging Horizon rover into this lab. Qdrant. LangSmith on StratifyLabs. Default README extras. --- ### Bias for Action — primary Horizon ERC date **Primary project:** Team Horizon Feb–Jun 2024 — ERC 2024 immovable date; ship 60 FPS + costmap on the hardware we had. **Backup project:** Eight hackathons (CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up) / Ylogx: measure bot DB latency, ship Redis, −35%. **Spoken STAR** Situation. From February to June 2024 I was core software for Team Horizon at CUSAT, building a semi-autonomous Mars rover for the European Rover Challenge 2024. More than eighty teams. The event date does not move because your Gazebo world is still ugly. Waiting for a perfect lab stack after the event is not a decision. Waiting for extra sensors as the first move is delay dressed as prudence. An irreversible collision is not the kind of decision you rush — speed without a live costmap is a blind rover. Task. Perception, mapping, and planning had to run on the hardware we had, in dirt and lighting that the simulator did not fully copy. I owned the software path that had to actually run on challenge day: a live camera feed, a dense map, and a costmap the planner could trust. Action. I shipped a real-time camera pipeline at 60 FPS with GStreamer instead of a naive Flask MJPEG stream that would have dropped the operator feed. Public repo `Gstreamer-UDP` is a webcam-stream artifact in the same family, not a second project. I took ZED 2 mapping at more than two million points per second into RViz and Gazebo so we had occupancy, not a pretty point cloud on a laptop. I put costmap planning and predictive sensor fusion in front of the actuators. Downsample for visualization, keep density for the local costmap so two million points per second did not all try to leave the rover. Field time was part of the action: Gazebo lighting was not dirt. I sequenced 60 FPS and the costmap first; extra hardware was not the first spend. Result. We placed 17th globally out of more than eighty teams. Obstacle detection accuracy improved 40 percent. Collision risk dropped 55 percent after costmap fusion, not after buying another sensor. The resume does not state a miss. I will not invent one. **Why-chain** 1. Why GStreamer at 60 FPS, not TCP MJPEG? The operator and autonomy path needed a pipeline, not a screenshot stream. Acting on a blind rover is the wrong bias. 2. Why costmap fusion, not wait for another sensor? The date does not wait on a purchase order. −55 percent collision risk is software on the sensors we had. 3. Why sequence 60 FPS and costmap before more mapping polish? A fast video that still collides is motion, not a result. Stale occupancy is the ship-blocker: operator sees 60 FPS; planner sees last second’s rocks. **BR follow-ups** 1. **What you did vs team.** I owned GStreamer, ZED mapping into the costmap, and fusion on the path to actuators. 17th is a team score. I will not claim I built the mechanical rover. 2. **Metric how-measured honesty.** 17th/80+, 60 FPS, 2M+ pts/s, +40% obstacle detection, −55% collision risk. I will not invent a p95 on the pipeline. Three clocks in debug: GStreamer pipeline latency versus ROS2 callback versus ZED rate. If I don’t remember a protocol: the resume numbers are those; the mechanism was live costmap plus fusion. 3. **What you’d do differently.** Time-box “one extra sensor versus fusion” against the collision-risk number earlier. Sequence 60 FPS and costmap on a calendar to the ERC date even earlier. Name the ship-blocker in week one, out loud. 4. **10×.** There are not 10× ERC teams to serve. Honest scale is 10× point-cloud rate or 10× field hours: you still cannot ship 2M+ pts/s to a laptop; downsample viz, keep local density; stricter time-sync. Not Kafka, not a cloud GPU bill, not Argus YOLO on this rover. 5. **Conflict.** Software costmap versus “buy another sensor.” We committed to fusion. Technical. No fight. 6. **Failure inside the same story.** Perception without planning — a fast video that still collides. Stale costmap. Sim lighting ≠ dirt. Reckless action would have been skipping the costmap to look busy. I did not. I also did not miss ERC; I will not invent that. **What you will NOT say.** Missed ERC. Haar. YOLO on the rover. A purchase-order drama. p95. “I was team captain.” Mentee count of juniors on the rover. --- ### Frugality — primary Ylogx Redis −35% + CloudFront/ECS **Primary project:** Ylogx — Redis on the bot path; CloudFront + ECS + Docker; GoDaddy → Route 53 → ALB; not a bigger RDS first. **Backup project:** Horizon costmap and fusion versus adding hardware; collision risk −55%. **Spoken STAR** Situation. Ylogx ran BI and SQL RAG on Postgres. Every natural-language question hitting the database is both latency and money. The obvious spend is a larger instance. Frugality is not cheapness. It is not spending on the wrong layer first. There are no extra points for budget size. Task. Cut database cost and latency on the hot bot path without skipping isolation, and keep dashboards off a foundation model. Action. I measured bot database latency instead of buying RDS headroom. I put Redis cache-aside on schema lookups and repeat answers; keys are tier-correct under RLS. I deployed CloudFront plus ECS plus Docker CI/CD instead of over-provisioning a snowflake host. I used one routing path — GoDaddy DNS into Route 53 into an ALB — and held sub-210 milliseconds on dashboards. Cached answers still had to be tier-correct. A dashboard KPI that must be sub-210 milliseconds does not go through an LLM at all. Playwright-every-URL thinking belongs to IQVIA; here the frugal move is cache plus pooling, not a bigger model. Result. Minus 35 percent bot database latency. Sub-210 milliseconds. 99.9 percent uptime. 40 percent faster reports. I will not invent an AWS bill or a TTL. **Why-chain** 1. Why Redis, not a bigger RDS? Repeat NL questions and schema metadata are cache-shaped. Minus 35 percent is the resume number. 2. Why still RLS after the cache? A cached answer for the wrong org is a leak. Frugality that leaks is not frugality. 3. Why CloudFront/ECS instead of a bigger host? Serving path, not a snowflake. Sub-210 ms is cache plus pooling, not an LLM on the hot path. **BR follow-ups** 1. **What you did vs team.** I put Redis on the bot path and kept dashboards off the LLM. I did not own the AWS bill. I will not invent one. 2. **Metric how-measured honesty.** −35%, sub-210 ms, 99.9%, 40%. No p95. No TTL. No dollar amount. If they ask cost: I don’t remember a bill; the resume number is minus 35 percent latency; the mechanism was Redis still behind RLS. 3. **What you’d do differently.** Write cache invalidation rules down with RLS from day one. Denied cross-tier cache keys as tests. 4. **10×.** More cache hits, still tier-scoped; pooling; CloudFront for static. Not a bigger RDS as the first 10× move. Not Kafka. 5. **Conflict.** Bigger instance versus cache. Technical. Backup: Horizon extra sensor versus fusion — that is Frugality on student hardware, not a fourth Ylogx primary. 6. **Failure inside the same story.** Uncached SQL RAG. The chatbot worked. The database paid for every turn. Instinct blamed the generator. **What you will NOT say.** AWS bill saved $N. TTL. p95. SEO. Qdrant. “I cut the company’s cloud spend by…” --- ### Earn Trust — primary GiftedBooks RAG + 99.5% **Primary project:** GiftedBooks — vocally self-critical that VR without grounded Q&A is decoration; 99.5% uptime; hours → 3–10 min. **Backup project:** Ylogx three-tier RLS: speak candidly that a JOIN leaks if isolation is only in the UI. **Spoken STAR** Situation. GiftedBooks is a VR learning suite: 3D labs, AI avatars, and PDFs. Official Earn Trust: listen, speak candidly, vocally self-critical, do not believe the work smells of perfume. Customer-product trust is also Customer Obsession. This LP is how honest I am about my own work. The awkward sentence I had to say out loud is that a pretty lab that cannot answer from the student’s own material is decoration, and a generic chatbot that sounds like a textbook is not a teaching assistant. Task. Say the failure mode before we dressed it up: hallucinated citations and downtime both burn trust. Then ship the honest product. Action. I will not pretend the VR surface was enough. I shipped contextual PDF Q&A, not a general LLM with a textbook voice. Retrieve from the upload; generate small; abstain if the answer is not in the chunks. I held sub-300 millisecond API responses and 99.5 percent uptime so the assistant was there during study. I used PYQ-prioritized topics so the product is honest about exams. Per-user PDFs must not retrieve another student’s notes. That isolation is trust, not a feature checkbox. Result. Doubt time from hours to three to ten minutes. Reading efficiency up 35 percent, engagement up 50 percent, comprehension 2.5 times. 99.5 percent uptime. I will not claim zero incidents. I will not say the body of work smelled of perfume because the VR looked expensive. GitHub AegisAI mismatch: resume product only. **Why-chain** 1. Why is this Earn Trust, not only Customer Obsession? Customer Obsession is start with the student. Earn Trust is saying the VR was not enough, out loud, and refusing off-PDF answers. 2. Why grounded RAG, not a fluent tutor voice? If the model answers a definition that is not in the PDF, the student memorizes a fluent lie. Citation from the chunk is the mechanism. 3. Why 99.5 percent as trust? An assistant that disappears mid-session trains people not to rely on it. I will not claim 99.9 percent; that number is Ylogx. **BR follow-ups** 1. **What you did vs team.** I named the decoration failure and shipped RAG + PYQ + uptime/latency bars. I did not claim I built every avatar. 2. **Metric how-measured honesty.** Same GiftedBooks resume numbers. No student headcount. 99.5% not 99.9%. No LangSmith claimed. 3. **What you’d do differently.** Add the same citation-eval habit as IQVIA. Not claimed on the resume — I would add it. 4. **10×.** Per-user isolation; embed once; small generate. Trust at 10× is “do not retrieve another student’s notes,” not a bigger model. 5. **Conflict.** Product backbone: no, we will not answer off-PDF. That is pushback, not a rude customer fight. See §3. 6. **Failure inside the same story.** Pretending the VR lab was enough. First pass without PYQ still left people guessing what to study. **What you will NOT say.** Zero hallucinations. Student headcount. Mixing AegisAI. Stealing Argus 73% as a fourth GiftedBooks/Argus mashup. Ylogx as a third GiftedBooks primary — Ylogx is backup only here so Ylogx does not take a fourth primary. --- ### Dive Deep — primary Ylogx Redis + RLS **Primary project:** Ylogx — anecdote said the chatbot worked; metrics said bot DB latency and cross-tier risk. Redis −35%. Three-tier RLS. Not SEO. **Backup project:** IQVIA — wrong chunks on 200+ page BRDs; hybrid plus semantic versus vector-only plus GraphDB plus LangSmith. **Spoken STAR** Situation. The Ylogx SQL RAG chatbot was “working” while bot database latency and cross-tier risk were the real product. On a happy-path account the model wrote SQL, Postgres returned rows, and a non-technical user nodded. That is the anecdote. Official Dive Deep: operate at all levels, stay connected to the details, audit frequently, be skeptical when metrics and anecdote differ, no task is beneath you. Measuring that path myself was not beneath the intern ticket. Task. Find whether slowness was query shape, missing cache, or security checks done in the wrong layer, and make three organizational tiers real on the path the bot and the reports already shared. Action. I measured bot database latency instead of celebrating that SQL came back, and instead of blaming the generator. That is a different question from “does the LLM write syntactically valid SQL.” I put row-level security and RBAC in Postgres for those three tiers so generated SQL ran as the user’s database role. A JOIN that is legal SQL and illegal for the tenant fails closed. I then put Redis on the hot bot path so repeat questions and schema lookups were not a round trip every time, without skipping RLS. A cached answer still had to belong to the right tier. Showing the SQL to the user is how wrong grain shows up even when the tenant is correct. Result. Bot database latency dropped 35 percent. Analysis productivity went up 65 percent. Report generation got 40 percent faster. 99.9 percent uptime held. Dashboards stayed in the sub-210 millisecond number. The deep dive was latency plus isolation. It was not a search-console ticket, and I will not lead with 403 or noindex. **Why-chain** 1. Why Redis not a bigger RDS? Repeat NL and schema lookups are cache-shaped. The hot path was not “RDS too small” as the first diagnosis. 2. Why still RLS after the cache? A cached leak is still a leak. App-only filters fail on a JOIN. 3. Why not start by rewriting the prompt? Assuming the generator was wrong failed first. The metrics disagreed with the demo anecdote. **BR follow-ups** 1. **What you did vs team.** I measured, I put RLS in Postgres, I put Redis on that path. Intern. Not beneath me to time the database. 2. **Metric how-measured honesty.** Minus 35 percent bot database latency is the number I will defend. I will not invent p95 or TTL. 99.9%, +65%, 40%, sub-210 ms sit on the same product. 3. **What you’d do differently.** Denied cross-tier RLS cases as tests from day one. Cache keys as tested artifacts with RLS. 4. **10×.** Cache hits still tier-scoped; pooling; RLS must hold. 10× NL questions without RLS is 10× leaks. 5. **Conflict.** Metrics versus anecdote inside the team: “it works in the demo.” I stayed on the measurement. No fight scene. Backup conflict is IQVIA retrieval if they already used Ylogx. 6. **Failure inside the same story.** Believing the demo. Generator-blaming. App-only isolation. **What you will NOT say.** SEO 403 / www versus non-www / noindex as the lead. p95. TTL. AWS bill. Qdrant. Django CUSAT portal (off-resume). Haar. --- ### Have Backbone; Disagree and Commit — primary IQVIA hybrid vs vector-only **Primary project:** IQVIA — hybrid + semantic Azure AI Search + GraphDB vs ship-vector-only; then commit on LangSmith evals. **Backup project:** Ylogx RLS-in-DB versus app-only filters; commit: three-tier RLS as platform rule. **Spoken STAR** Situation. The default for “RAG on PDFs” is a vector index and a demo. IQVIA BRDs are more than two hundred pages. Test-case generation from the wrong section is expensive. A teammate or stakeholder who wants vector-only is not being foolish; they are optimizing for a ship date. Official Backbone: respectfully challenge; do not compromise for social cohesion; once a decision is determined, commit wholly. I do not have a named interpersonal blow-up on the resume. Task. State the failure mode out loud without turning it into a personality fight. Then, if we choose a path, instrument it instead of keeping a quiet side project. Action. I stated that similar embeddings are not requirement coverage. Clause IDs and tables miss on cosine. I proposed Hybrid plus Semantic Azure AI Search plus a GraphDB plus LangGraph adaptive retrieval, with LangSmith traces so we could see a bad path. Lexical for “R-141,” semantic for policy language, graph hops for dependencies, cite or abstain. Once we chose that path, I committed by instrumenting evals and test-case generation instead of reopening the tool debate every week. Commit wholly means trace the chosen path, not keep a quiet FAISS side project. I did not overrule a named manager. I will not perform a fight. Result. Hybrid RAG as on the resume. Two-hundred-page BRDs. Evals and test-case generation. Disagree on retrieval. Commit on evals. **Why-chain** 1. Why challenge vector-only at all? Coverage. “R-141 must be retrievable on Monday” versus “ship the vector demo by Friday.” The disagreement is retrieval failure mode, not personality. 2. Why commit on evals, not on winning the tool argument? Once the decision is determined, Backbone is instrument the path. Reopening libraries every week is not commit. 3. Why LangSmith, not a feeling that hybrid is better? Fail if a generated test case has no overlapping span. That is how commit becomes visible. **BR follow-ups** 1. **What you did vs team.** I argued the failure mode and I instrumented evals. I did not overrule a named manager. There is no quote. 2. **Metric how-measured honesty.** 200+ page BRDs processed with traces and test-case gen. No extra percent. Success is not fluency. 3. **What you’d do differently.** Put the eval table on the table *before* the argument so commit is on numbers. Freeze the eval slice earlier. 4. **10×.** Evals so wrong answers do not scale across 10× documents. Hybrid retrieve small chunks. Not a bigger embedding model as the first 10×. 5. **Conflict.** This *is* the conflict card. Technical. Stay on the failure mode if they push “what did the other person say.” I do not have a transcript. 6. **Failure inside the same story.** Vector-only on long BRDs. Reopening the tool debate instead of tracing — that would have been fail-to-commit. I traced. **What you will NOT say.** Manager fight. Raised voice. “I overruled my manager.” Qdrant. A name. Fake intern drama. --- ### Deliver Results — primary Horizon 17th/80+ **Primary project:** Horizon ERC 2024 — 17th of 80+; 60 FPS; costmap; +40% obstacle; −55% collision. Output, not activity. **Backup project:** Ylogx 40% faster reports, 99.9% uptime, sub-210 ms — use if intro already used the rover. **Spoken STAR** Situation. ERC 2024 had more than eighty international teams. I was Horizon software from February to June 2024. Official Deliver Results: key inputs, right quality, timely fashion; despite setbacks, never settle. Activity would have been more bags and more Gazebo screenshots. The result the judges score is a rover that runs the challenge on the day. The setback was time-sync: the operator feed could look live while occupancy was already old, and Gazebo lighting was not dirt. Task. Ship a live feed plus a costmap the planner could trust, on the date, on the hardware we had. A lab video is not a result. A rover that maps in bags and still collides is not a result either. Action. I shipped GStreamer at 60 FPS. I took ZED 2 mapping at more than two million points per second into RViz and Gazebo. I put costmap planning and sensor fusion in front of the actuators. When the costmap lagged, the rover planned on stale occupancy. That was the ship-blocker, and I treated it as one, not as someone else’s camera ticket. We downsampled for visualization and kept density for the local costmap. Perception, mapping, and planning stayed separate ROS2 nodes. mAP belongs to Argus; I will not mix stacks. Result. We placed 17th globally of more than eighty teams. Obstacle detection accuracy improved 40 percent. Collision risk dropped 55 percent after fusion, not after a purchase order. The date was met because 60 FPS and a live costmap were sequenced first. The resume does not state a miss. I will not invent one. **Why-chain** 1. Why is 17th a result and more bags not? Judges score the run. Perception without planning is activity. 2. Why fusion over extra hardware as the path to the result? Student hardware, immovable date. −55% is the quality metric on the result. 3. Why talk stale costmap if you placed 17th? Because that was the setback inside the win. Bar Raisers want the defect, not the trophy. **BR follow-ups** 1. **What you did vs team.** Software path: GStreamer, ZED, costmap, fusion. 17th is a team score. Mechanical/electrical were teammates. 2. **Metric how-measured honesty.** 17th/80+, +40%, −55%, 60 FPS, 2M+ pts/s. I will not invent a missed ERC to sound humble. If intro already used this, I switch to Ylogx 40%/99.9% as Deliver Results backup — a fast leak is not a delivered result; RLS was part of that result. 3. **What you’d do differently.** Sequence demo-risk items on a calendar to the ERC date even earlier. Time-box extra-sensor versus fusion against collision risk earlier. 4. **10×.** Same as Bias: point-cloud rate / field hours, downsample viz. Not Kafka. 5. **Conflict.** Extra sensor versus fusion. Technical. 6. **Failure inside the same story.** Stale occupancy. Sim ≠ dirt. Fast camera with no planning looks like motion and still collides. **What you will NOT say.** Missed ERC. Cut on 99.9% (wrong product). Haar. YOLO on rover. “I carried the team.” --- ### Strive to be Earth’s Best Employer — primary Horizon knowledge sharing **Primary project:** Horizon — pair on launch files so camera/mapping knowledge is not in one head. Teammate environment, not Amazon HR. **Backup project:** IEDC CUSAT Tech Team. CGPA 8.42 is not this LP’s metric. **Spoken STAR** Situation. Horizon rover software is unusable if camera and mapping knowledge lives in one head. ERC is a team score. Official Earth’s Best Employer: safer, more productive environment; are teammates growing, empowered, ready for what’s next. Honesty: this is teammate environment on a student rover team, not Amazon HR. No retention survey. I was a software team member, not a people manager, not HR, and not a hiring bar raiser. Task. Growing means they can launch perception without me. Empowered means they are not blocked on my laptop. Ready for what’s next means a teammate can run the build at the next event or the next subsystem. Do not collapse this into Hire and Develop — here the bar is growth and ready-for-what’s-next, not “I coached / I raised a hiring bar.” Action. I worked as a software team member, not a hero module. I shared GStreamer pipeline constraints, ZED rate versus costmap, and launch order so others could integrate. Pair on launch files so a second person can start the camera node. Downsample/viz versus local density is knowledge that has to leave my notebook. I used the same habit at IEDC and at eight hackathons: the person next to you can run the build. I did not run a mentoring program with a name. Result. 17th of more than eighty is a team result. IEDC Tech Team and eight hackathons are on the resume. I will not claim a happiness score. Obstacle detection +40 percent and collision −55 percent are product metrics from the stack we could run together, not an HR survey. **Why-chain** 1. Why launch-file pairing, not a slide deck? Empowered means they can run it. A deck does not start GStreamer. 2. Why not call this Hire and Develop? Hire and Develop primary is IEDC bar-raising without a headcount. This LP is “are they ready for what’s next” on the rover. If they already heard IEDC, stay on Horizon launch files. 3. Why is 17th the result for this LP? Because a solo notebook does not place. The environment is what made the score possible. **BR follow-ups** 1. **What you did vs team.** Shared constraints, paired on launches. I did not manage people. I did not run DEI or parental-leave programs; I will not fake Amazon-scale HR. 2. **Metric how-measured honesty.** 17th/80+ as team result. No retention number. No mentee count. 3. **What you’d do differently.** Pair earlier on launch files. Ten-line runbook (not claimed as a wiki). 4. **10×.** Not 10× employees. Honest: more subsystems launchable by more people on a student team. 5. **Conflict.** Hero-module versus shared launches. I chose share. No fight. 6. **Failure inside the same story.** Siloed ROS nodes: one person could run perception and nobody else could launch planning. **What you will NOT say.** Mentee of 12. Happiness score. “I was their manager.” Amazon DEI program I did not run. Fake people-management. --- ### Success and Scale Bring Broad Responsibility — primary Argus 20+ cameras after 89% **Primary project:** Argus — do not multiply a 73% model across 20+ cameras; alarm fatigue is the secondary effect. **Backup project:** Ylogx ALB + ECS + CloudFront: scale the BI path that already has RLS. 99.9%, 210 ms. Scaling SQL RAG without cache and RLS scales leaks. **Spoken STAR** Situation. One camera demo does not create industrial safety. Wrong alerts at more than twenty cameras scale harm as well as good. Alarm fatigue is the downstream effect. Official Success and Scale: thoughtful about secondary effects; leave things better. I will not invent a planet metric. On this resume the secondary effect is alarm fatigue if you scale a weak detector. “Leave things better” is this LP, not Ownership. Task. Get the eval right, then multiply. Do not look busy by attaching twenty streams to a 73 percent model. Action. I did not scale 73 percent mAP. I got to 89 percent mAP at 24 FPS on more than 15,000 images, with NMS so duplicate boxes do not become duplicate alarms. Then I containerized more than twenty feeds, wrote Postgres logs, and shipped alerts. Video stays off the database; events go in. I will not invent EKS because Ylogx is the ECS story. Result. More than twenty cameras, minus 50 percent violations, 2 times compliance, after the eval gate, not before it. I will not claim a climate or community program. **Why-chain** 1. Why not scale first and fine-tune later? Secondary effect: alarm fatigue trains people to ignore PPE. That is harm at 20×. 2. Why 24 FPS at 20+ cameras as a scale constraint? A correct box that cannot keep up is a silent miss multiplied. 3. Why Postgres events as part of leaving things better? Audit trail. The next person can see what fired. Video blobs in the database would not leave the ops path better. **BR follow-ups** 1. **What you did vs team.** Gate, then containerized path, then logs. I will not invent a 20-site deployment I did not run. 2. **Metric how-measured honesty.** 20+, 89% mAP, 24 FPS, −50%, 2×. After the gate. Same numbers as Are Right / Highest Standards — if they already heard those, start on *secondary effects* and 20× cameras, do not recap the training curve as if it were new. 3. **What you’d do differently.** Per-camera error budget. The resume does not state it; I would add it. I am not claiming I shipped it. 4. **10× cameras.** Workers per stream or batch; do not run 20 copies of the fattest model if 24 FPS breaks; mAP gate then multiply. Backup Ylogx: 10× NL without RLS scales leaks; ALB/ECS with RLS is the serving scale I will defend. Not a fourth Ylogx primary. 5. **Conflict.** “Go live on twenty cameras this month” versus the gate. Technical. I would not multiply 73 percent. 6. **Failure inside the same story.** Scaling a 73 percent model. That would have left the floor worse. I stopped at the gate first. **What you will NOT say.** Planet / climate metric. EKS. Community program. 95 mAP. Haar. Kafka. --- ## 2 Genuine stories — second example for the LPs Bar Raisers re-ask Bar Raisers often demand a SECOND example. These are genuine and on a **different project** than the primary. If they already used the second project as someone else’s primary in this loop, go to the defect inside the remaining story — do not invent a third company. Do not break the primary counts by promoting these into a fourth primary for Ylogx / IQVIA / Argus / Horizon. These are **seconds when asked**, not a new spread. --- ### Ownership — second: Horizon GStreamer + mapping beyond “just camera” **Different from primary Ylogx.** Situation. On Horizon I was software team, not “the camera person” on an org chart that did not exist. A webcam-quality demo stream was the designated work for a lab video. ERC 2024 still needed a feed and a map the planner could trust. Task. Own the ship-blockers even if they were outside a narrow camera ticket. Never “that’s not my job” when the rover is blind or the costmap is stale. Action. I shipped GStreamer at 60 FPS on the operator and autonomy path. I took ZED 2 at more than two million points per second into a costmap, not a pretty cloud. Fusion in front of actuators. Downsample viz, dense local costmap. Result. 17th/80+, +40% obstacle detection, −55% collision risk. Team score; I owned that software path. If they ask what I did versus team: I did not own mechanical. I owned the pipeline that had to run. Differently: name the ship-blocker in week one. 10×: point-cloud handling, not Kafka. Conflict: extra sensor versus software — we fused. Failure: stale occupancy while the feed looked live. --- ### Dive Deep — second: IQVIA retrieval + ranking contradiction **Different from primary Ylogx. Do not lead with SEO.** Situation. Hybrid RAG on more than two-hundred-page BRDs. Anecdote: the demo retrieved a similar paragraph. Metrics/evals: clause IDs missed; two high-ranked pages disagreed; first-hit retrieval looked confident. Task. Stay connected to chunks, traces, and agreement, not to fluency. No task beneath me: read the retrieved span, not only the writer’s paragraph. Action. Hybrid lexical + semantic on Azure AI Search. GraphDB hops for dependencies. LangGraph adaptive retrieval. Rank Deep Research by quality, recency, and **agreement**. LangSmith: empty scrape still feeding the writer; fail if no overlapping span, including generated test cases. Citation-gated writer. Result. 200+ page BRDs processed with traces; 200+ sites ranked on the research side. No extra percent. Why Redis is not this card: that was Ylogx. Why hybrid: cosine misses “R-141.” Why agreement ranking: two pages disagree; first-hit is false certainty. Differently: freeze gold questions before Playwright. 10×: cache scrapes, small chunk retrieve. Conflict: vector-only versus hybrid — Backbone primary; here stay on the *details* of the wrong chunk. Failure: stuffing the PDF; linear chain; first-hit confidence. --- ### Deliver Results — second: Ylogx 40% / 99.9% **Different from primary Horizon. Use this if they already heard the rover.** Situation. At Ylogx the result that mattered was not “the chatbot demoed.” Operators needed custom reports, NL over warehouse facts, and live KPI tiles — without one tenant reading another. Task. Deliver faster reports at held uptime, with non-technical users actually able to ask questions. A fast leak is not a delivered result. Action. FastAPI plus NestJS plus Postgres. Three-tier RLS so generated SQL ran as the user role. Redis on the bot path, minus 35 percent database latency. CloudFront and ECS so dashboards stayed sub-210 milliseconds without putting an LLM on that hot path. Thirty KPI dashboards. Result. Reports 40 percent faster at 99.9 percent uptime. +65% analysis productivity. +60% ops efficiency on those dashboards. Isolation was part of the result. I vs team: I owned the data path and cache; I was not the named SRE org. Measured: resume numbers; no p95 invented. Differently: denied RLS tests day one. 10×: cache + RLS or you scale leaks. Conflict: app-only versus RLS-in-DB. Failure: demo “worked,” database paid, JOIN could leak. Do not invent a miss on 99.9%. --- ### Backbone — second: Ylogx RLS-in-DB vs app-only **Different from primary IQVIA.** Situation. The chatbot returned rows on a demo. App-only filters looked enough. A JOIN or generated query could outrun a UI filter and return another organization’s rows. Task. Disagree without a fight scene: isolation belongs in Postgres, not only in NestJS middleware. Then commit wholly to three-tier RLS as the platform rule. Action. I said I would not ship the bot on app-only checks. Generated SQL as the user’s database role. RBAC for three org tiers on the path the bot and reports share. Redis after, not instead: cache still tier-correct. Once that was the path, I did not keep a quiet “filter in the API” side door. Result. 99.9% held, −35% bot latency, +65% analysis productivity, 40% faster reports. Commit is the policies in the database, not a Slack argument I will invent. What the other person said: I do not have a transcript. Stay on the JOIN leak. Differently: eval table of denied cross-tier queries before the argument. No manager name. --- ### Learn — second: Horizon ROS2 / GStreamer / ZED **Different from primary IQVIA.** Situation. Web intern skills (FastAPI/NestJS later at Ylogx; this was 2024) were not ROS2, GStreamer clocks, or ZED 2 at more than two million points per second. ERC still had a date. Task. Learn enough to ship a live feed and a costmap, with a stop condition: the planner trusts occupancy, the operator sees 60 FPS. Not wandering in Gazebo screenshots. Action. I learned GStreamer pipelines instead of Flask MJPEG. I learned to take ZED density into a costmap, downsample for viz. I learned three clocks: pipeline latency, ROS2 callback, ZED rate. Field time: Gazebo lighting was not dirt. Curiosity stop: fusion on the sensors we had, not another week of “maybe another sensor.” Result. 17th/80+, +40% obstacle, −55% collision, 60 FPS, 2M+ pts/s. Differently: time-box sensor-versus-fusion earlier. Backup learn if they also used Horizon: eight hackathons, rapid production-ready deploy — I will not invent hours-to-prize. I will not steal IQVIA LangGraph into this paragraph. --- ### Are Right — second: Ylogx measured latency (anecdote vs metric) **Different from primary Argus.** Situation. Instinct said the SQL RAG generator was slow. The demo returned rows. The metric said every natural-language turn was hitting Postgres, and isolation in the wrong layer was a second way to be wrong. Task. Disconfirm “the model is thinking.” Be right about the layer: database latency and RLS, not a bigger prompt. Action. I measured bot database latency. I put RLS in Postgres. I put Redis on repeat NL and schema lookups. I did not buy a bigger RDS first. I did not treat syntactically valid SQL as “right” if grain was wrong — show the SQL. Result. −35% bot DB latency, 99.9% uptime, 40% faster reports, +65% productivity. Right is the measurement, not the demo. Differently: denied-tier tests day one. Do not invent p95. Failure: believing the anecdote. This is Are Right on measurement; Dive Deep primary is the same product — if they already used Dive Deep, start on “I was wrong that the LLM was the bottleneck,” not a second recitation of Redis. --- ### Bias for Action — second: Ylogx Redis ship (not wait for a rewrite) **Different from primary Horizon.** Alternate second if they already used Ylogx this loop: eight hackathons, CodeRecet 1st — shortest clock, production-ready deploy in the window; I will not invent hours. Situation. Bot path was paying Postgres on every turn. A rewrite of the whole BI stack would have been delay dressed as prudence. Calculated risk: cache-aside still behind RLS, not skip isolation to look fast. Task. Measure, ship Redis, keep dashboards off the LLM. Do not wait for a perfect schema redesign after users already ask NL questions. Action. I measured bot database latency. I shipped Redis on schema and repeat answers with tier-correct keys. CloudFront/ECS path already there; I did not wait to re-platform. I refused to put the sub-210 ms dashboard on an LLM to look like action. Result. −35% bot DB latency. 99.9% held. 40% faster reports. Reckless action would have been skipping RLS. I did not. Differently: invalidation rules written with RLS on day one. Hackathon alternate: eight events, first at CodeRecet, MLH Best Project, Magnathon 2.0 runner-up — rapid software development and production-ready deployment as on the resume; no invented prize clock. --- ## 3 Exact IE prompts (map, don’t invent 10454435) Every prompt below was asked in some other SDE I / UTA / AUTA / India SDE I later live (HM / BR / R3). Sources are in `interview_notebook/R3-R4-Question-Bible.md` and the parent list. Job **10454435** still names **none**. Unnamed stays unnamed. Speak; do not dump LP titles. --- ### Tell me about a time you failed to deliver the expected result **Map.** Rudraksh BR “mistake”; GFG 2025 / Rudraksh / LC 6653463 negative feedback; Reddit 1s9z862 Failure (Job 3100855, not 10454435). Honest on-resume: Argus 73% mAP was not the expected safety result. Do **not** invent missed ERC or missed 99.9%. **Spoken.** The expected result on Argus was a PPE detector I would actually alert a human on. 73 percent mAP was not that. At that number you miss helmets and you fire junk boxes. Either one trains people to ignore alarms. I treated a notebook curve as progress. That failed the expected result. I did not invent a manager quote. I trained on more than 15,000 images, iterated YOLOv9 to 89 percent mAP, required 24 FPS, NMS on boxes, Postgres event logs, then more than twenty cameras. Violations down 50 percent, compliance doubled, after the gate. I will not invent a missed ERC or a missed 99.9 percent to make this sound like a calendar miss. If they want a second failure-to-deliver: Ylogx chatbot “delivered” rows and failed to deliver isolation and latency until RLS and Redis. STOP. If they 10×: do not scale 73 percent. --- ### Took extra responsibility / went beyond scope / issue not your task / outside designated work **Map.** IE.in 2024-grad AUTA; LC 7850431; LC 6570344; GFG sde-1-17; LC 6806195 “went beyond task”; LC 8362604 “problem I identified myself.” **Spoken — default Ylogx RLS as platform, not chatbot-ticket.** On paper my Ylogx ticket was a chatbot and a report builder. Isolation and bot database latency were not “just SQL generation.” A chatbot that leaks a tier or waits on Postgres is not a chatbot you can give to a non-technical user. I found that while delivering the feature I was asked for. I put three-tier RLS and RBAC in Postgres so generated SQL ran as the user’s role. I measured bot database latency and shipped Redis for minus 35 percent without skipping RLS. 99.9 percent held. +65 percent analysis productivity. 40 percent faster reports. Platform isolation was not the demo ticket. The UI filter was the tempting ticket-sized fix. It was the wrong layer. **If they already heard Ylogx — Horizon.** I was software, not “just camera.” GStreamer at 60 FPS and ZED into a costmap were outside a lab-video stream and inside ERC. 17th/80+, +40%, −55%. A Flask MJPEG was in scope for a screenshot. 60 FPS was outside that screenshot. **IQVIA spare.** Test-case generation from BRDs expanded RAG into something a human could audit, LangSmith traces, 200+ page documents. --- ### Conflict with manager or teammate **Map.** GFG 2025 BR; GFG April 2026 R3/HM; GFG sde-1-17 client POV; LC 8014509 / 7724048 “convince someone”; Reddit 1s9z862 disagreement with senior (different job id). **Spoken.** I do not have a named interpersonal blow-up on the resume, and I will not invent a manager quote or a raised-voice scene. The conflict I will speak is retrieval at IQVIA. The fast path is a vector index and a demo. A teammate or stakeholder who wants that path is optimizing for a ship date. Two-hundred-page BRDs have clause IDs cosine misses. I stated similar embeddings are not coverage. I argued for hybrid plus semantic Azure AI Search plus GraphDB plus LangGraph adaptive retrieval, with LangSmith traces. Once we chose the path, I committed on evals and test-case generation instead of reopening tools every week. Disagree on retrieval. Commit on evals. **Backup.** Ylogx: RLS-in-DB versus app-only. JOIN leaks. Commit: three-tier RLS as platform. Optional Horizon: costmap versus extra sensor; −55% collision. If they push “what did they say”: I do not have a transcript. The disagreement was coverage versus ship-date, or JOIN leak versus UI filter. I did not overrule a named manager. --- ### Ownership & accountability; collaboration; challenges; decision-making; impact; learning from failures **Map.** Uday BR collaboration/ownership UNNAMED; Ansuman BR decision-making/ownership UNNAMED; Kushank HM projects/ownership; Aditya trio; Nisarg all-16; LC 7981646 resume metrics/ownership. Exact wording UNNAMED on many. Job 10454435 still none. **Spoken — six beats, three projects, then stop. Do not recite sixteen names.** Ownership and accountability: Ylogx. I owned three-tier RLS on the bot and report path as an intern. Generated SQL as the user role. 99.9% held, 40% faster reports, −35% bot latency. Accountability is the database refused the wrong row, not a UI hide. Collaboration: Horizon / IEDC. I shared GStreamer, ZED, costmap launch order so the team could run the stack. 17th of more than eighty is a team score. I am not a manager. No mentee count. Challenges: IQVIA Hybrid RAG on 200+ page BRDs, or Horizon stale costmap — pick **one** if they already heard the other. Hybrid plus graph plus LangSmith stop; or 60 FPS plus live costmap. Mechanisms, not a list. Decision-making: IQVIA hybrid versus vector-only, then commit on evals. Or Horizon fusion versus extra sensor. Discarded option named. Impact: pick the metric on the story you are in. Do not dump every resume number. Learning from failures: Argus 73% mAP. Gate at 89% and 24 FPS before 20+ cameras. −50% violations, 2× compliance after. I will not invent a missed SLA. If they want Aditya’s three named prompts: Ownership Ylogx, pressure Horizon, learn IQVIA. Three projects. --- ### Implement something challenging / significant technical challenge / challenging implementation **Map.** GFG April 2026 R3/HM; IE.in 2024-grad AUTA BR; Raghav “most difficult”; LC 8362604 BR “challenge I solved”; LC 6475219 project critical work. **Spoken — pick one.** IQVIA Hybrid RAG on more than two-hundred-page BRDs. Stuffing the PDF invents citations. Vector-only misses clause IDs. A linear chain cannot retry or hop a requirement graph. I built hybrid plus semantic Azure AI Search, GraphDB hops, LangGraph adaptive retrieval, LangSmith traces, citation-gated writer, test-case generation as a traced job. Lexical for IDs, semantic for policy, graph for dependencies. 200+ page BRDs. 200+ sites on Deep Research. No extra percent. No Qdrant. Physics alternate: ZED 2 at 2M+ pts/s plus a costmap that must not go stale, 60 FPS, 17th/80+. Eval alternate: Argus 73→89 at 24 FPS on 15k+ images. I pick one and go deep. --- ### Critical feedback / last negative feedback / tough feedback **Map.** GFG 2025 BR; Rudraksh BR; LC 6653463; LC 6806195; LC 7850431; IE.in 2024-grad; Dhananjai BR. **Spoken.** The honest on-resume version is Argus, and I will say that if they want a person’s name I do not have. 73 percent mAP was the feedback. It was not production quality for PPE. Misses and junk boxes. I did not treat 73 percent as a blog metric. 15,000+ images, YOLOv9 to 89 percent mAP, 24 FPS, NMS, Postgres logs, then 20+ cameras. Violations −50%, compliance 2×, after the gate. I do not have a confirmed manager quote. I will not invent one. Backup if they already heard Argus: the system’s feedback on Ylogx was bot database latency before Redis, −35%. --- ### Deep dive into subject matter / last time you deep-dived a bug / deep dive to find a problem **Map.** GFG 2025 named Dive Deep; LC 7850431; LC 7724048; LC 8014509; IE.in 2024-grad; LC 6653463. **Spoken.** Ylogx, not SEO. Anecdote: chatbot returned SQL and rows. Metrics: every NL turn hit Postgres; JOIN could leak across orgs. I measured bot database latency. RLS + RBAC in Postgres for three tiers; generated SQL as user role. Redis on the hot path, −35% database latency, cache still tier-correct. 99.9% held. +65% analysis productivity. 40% faster reports. Sub-210 ms dashboards. Deep dive was latency plus isolation. If they already heard Ylogx: IQVIA ranking contradiction — two high-ranked pages disagree; first-hit is false certainty; agreement plus cite-or-abstain; LangSmith empty scrape feeding the writer. Or Argus 73% as an eval bug. Or Horizon stale costmap: operator 60 FPS, planner on last second’s rocks. SEO 403 is prep-only. --- ### Didn’t know how to proceed / options considered / limited information / comfort zone / stuck **Map.** LC 7850431 didn’t know what to do next; LC 7280347; LC 7406809; GFG intern stuck; GFG April 2026 limited information; LC 6573582 comfort zone. **Spoken.** The first time I had to architect multi-agent research plus hybrid retrieval on two-hundred-page docs, I did not know the next node. Ylogx had been FastAPI, NestJS, SQL RAG. A tutorial was not a graph. Options I considered: one scraper versus Firecrawl + Bing + DuckDuckGo + Playwright in parallel with LangGraph checkpoints; one vector index versus hybrid lexical+semantic Azure AI Search plus GraphDB hops; linear chain versus graph; stuffing the PDF versus retrieve-and-cite; fluency versus LangSmith fail-if-no-span. I discarded Qdrant; resume is Azure. I discarded first-hit ranking when two pages disagree. Stop condition: no overlapping span, including generated test cases. Result: 200+ sites ranked, 200+ page BRDs. Limited information is the job on Deep Research: no single gold page; rank by agreement; refuse without a span. Horizon backup: Gazebo did not tell us dust or lighting. Options: wait for another sensor versus fuse ZED 2 into a live costmap. We fused. −55% collision, 17th/80+. 60 FPS so the operator saw the same incomplete world in time. Who I consult when blocked: traces and evals first (LangSmith spans, mAP, measured latency), then teammates on the node I do not own (Horizon launches, IEDC pair). I will not invent a named mentor program. I do not escalate a JOIN leak to “someone else’s security team” and walk away — that is Ownership. --- ### Improved upon shortcomings **Map.** Neighbor of mistake / Highest Standards / LC 8362604 simplifying a process; “improved upon” wording appears in generic later-slot banks. Map to genuine shortcoming, not a fake SLA recovery. **Spoken — Argus shortcoming, then gate.** The shortcoming was 73 percent mAP. I improved it by data (15k+ images), YOLOv9 iteration, NMS, 24 FPS as a second bar, then scale. −50% violations, 2× compliance after. **Ylogx shortcoming.** SQL RAG without cache and with isolation in the wrong layer. Redis −35%, RLS in DB, 99.9% held. **IQVIA shortcoming.** Vector-only / no span check. Hybrid + GraphDB + LangSmith stop. **Simplifying a process (same family).** IQVIA test-case generation from BRDs so humans are not handwriting cases from 200+ pages. StratifyLabs browser inference, −30% ML iteration, versus a GPU loop per tweak. GiftedBooks PYQ instead of full-PDF grind, +35% reading. Redis versus bigger database, −35%. Simplifying without a mechanism is a deleted feature. --- ### Missed a deadline / what could you have done better / missed a commitment / struggled to meet a deadline **Map.** LC 8362604 BR; LC 7563011 R2; LC 7724048 BR; LC 6475219; Dhananjai; IE.in 2024-grad missed commitment. **Spoken. Lock.** I do not have a missed ERC or a missed 99.9 percent on the resume, and I will not invent one to sound humble. Amazon asks this because they want how you flag risk. ERC 2024 was fixed. More than eighty teams show up. I sequenced 60 FPS and the costmap so the run happened. 17th/80+. Collision −55%. Obstacle +40%. The alternate I will speak is sequencing: 60 FPS versus costmap — a fast camera with a stale grid is the miss that still looks like motion. I treated stale occupancy as the ship-blocker. Communication if a ship-blocker is late is early and specific: the feed is late, or the costmap is stale, or RLS is not in the database yet. Then we cut extra hardware or extra features. We do not cut isolation. We do not scale a 73 percent safety model. We do not skip the occupancy grid to look busy. What I could have done better: name that ship-blocker in week one, out loud; time-box extra-sensor versus fusion against collision risk earlier; calendar 60 FPS and costmap to the ERC date even earlier. Hiding a stale costmap until the last week would have been the miss. If they keep pushing for a personal miss: Argus 73 percent mAP is the honest “I was wrong about a ship bar,” not a calendar miss. --- ### Went beyond scope Same spine as extra responsibility. Speak Ylogx RLS as platform **or** Horizon GStreamer/mapping beyond just camera. Do not tell both at full length unless they ask for two. IQVIA test-case gen is the third spare. --- ### Complex project, what was complex, what would you do differently / POCs / multiple solutions **Map.** LC 7563011 AUTA HM (closest true-R3-after-two-lives shape); LC 7981646 Innovate/Frugality architecture. **Spoken.** IQVIA Deep Research and Hybrid RAG were the POC series. What was complex: one scraper is not enough — Firecrawl, Bing, DuckDuckGo, Playwright fail on different sites, so the graph fans out and checkpoints state. Vector-only is not enough on 200+ page BRDs, so the POC was hybrid plus semantic Azure AI Search plus GraphDB versus a single index. A linear LangChain chain cannot retry or hop a requirement graph. Stuffing the PDF dies on tokens and invented citations. Two high-ranked pages contradict; first-hit is false certainty. Evals in LangSmith were the stop: if a generated test case has no overlapping span, the path is not done. I discarded Qdrant. Result: 200+ sites ranked, 200+ page BRDs with traces and test-case generation. No extra percent. What I would do differently: freeze an eval slice of BRD sections to expected test cases before adding a fourth search tool; freeze gold questions with expected citations before Playwright; citation-required writer in the type system earlier; fewer Playwright renders, Bing snippets first when the snippet is thin. Backup POC: Ylogx FastAPI + NestJS + Postgres + Redis + RLS versus “just LangChain” and versus app-only filters. Discarded: service-role LLM writing SQL. --- ### Pushed back on a customer request **Map.** Neighbor of Backbone / Customer Obsession / GFG sde-1-17 extra for customers. If no genuine rude-customer fight: GiftedBooks grounded RAG is “no, we will not answer off-PDF.” Product backbone, not a fight. **Spoken.** I do not have a named angry-customer incident on the resume. The pushback I will defend is product. On GiftedBooks, a student wants an answer even when it is not in their PDF. Answering off-PDF is a fluent lie they will memorize. I pushed back by design: retrieve from the upload; if there is no overlapping span, abstain; keep generate small; PYQ ranker is analytics, not an LLM inventing the syllabus. Doubts hours → 3–10 min, sub-300 ms, 99.5% uptime, +35% reading, +50% engagement, 2.5× comprehension — on *grounded* help, not on unlimited hallucination. Ylogx spare: a user (or a demo stakeholder) wants “just show the rows.” Pushback: not without the three-tier RLS role. A JOIN without RLS is not customer service. IQVIA spare: analyst wants a fluent synthesis when two sources disagree. Pushback: rank by agreement; cite or abstain; LangSmith stop. I will not invent a shouting match. --- ### Designed with long-term vision / sacrificed short-term for long-term **Map.** Ownership official text (long-term value over short-term results); Think Big; Success and Scale. Exact “long-term vision” wording UNNAMED in several BR dumps. **Spoken — long-term vision, StratifyLabs primary Think Big.** Local GPU loop is short-term demo. Marketplace of 50+ models/datasets, browser inference, URDF/WebGL so the default path is reusable, Gemini RAG bots conditioned on the scene. −30% ML iteration. Long-term is reuse, not a fake TAM. I will not invent users. **Sacrificed short-term for long-term — Ylogx RLS (Ownership) or Argus gate (Success and Scale).** Short-term: ship the chatbot on app-only filters this week. Long-term: three-tier RLS in Postgres on the shared path, Redis still behind RLS, 99.9% held. Short-term Argus: twenty cameras this month at 73% mAP. Long-term: 89% and 24 FPS first, then scale; −50% violations after, not alarm fatigue. Horizon: extra bags this week versus a live costmap that still works on the ERC date. I did not cut isolation to make a date. I did not scale 73% to look busy. Those are the sacrifices I will not make. --- ### Feature customers didn’t explicitly request **Map.** Neighbor of Invent / Customer Obsession. Parent lock: Ylogx RLS; GiftedBooks PYQ ranker. **Spoken — pick one, go deep.** Ylogx RLS. Customers asked for a chatbot and reports. They did not ask for “Postgres row-level security policies.” They asked not to see another org’s KPIs, which is the same need in product language. I put RLS + RBAC for three tiers in the database because app-only is not that feature. 99.9%, +65%, 40% faster reports, −35% with Redis still tier-correct. GiftedBooks PYQ topic ranker. Students asked for answers and labs. They did not ask for “previous-year-question analytics as a separate job from the LLM.” They asked to pass exams. PYQ ranks topics from what exams actually ask. +35% reading efficiency. Doubts 3–10 min still need grounded RAG; PYQ is the study-path feature they did not name. Do not tell both at full length unless they ask for two. --- ### How do you approach a complex problem / who do you consult when blocked **Map.** LC 7563011 POCs; GFG April 2026 limited information; Rushikesh BR why / differently / measure; igreaper multi-level why. **Spoken.** I split the problem into a measurable failure mode, two or three POCs, and a stop condition. IQVIA is the picture: scrapers in parallel with LangGraph checkpoints; hybrid versus vector; LangSmith overlapping-span as stop. I do not add a fourth tool until the eval slice is frozen. Ylogx picture: measure bot DB latency; isolate in Postgres; cache without skipping RLS. Horizon picture: three clocks (GStreamer, ROS2, ZED); stale costmap is the failure mode; fusion on the sensors we have. Who I consult: traces and numbers first; then the person who owns the adjacent node (launch files, BRD consumer, dashboard). I do not invent a skip-level manager story. When blocked on retrieval, I read the span. When blocked on PPE, I look at mAP and fail cases, not an LLM description of the frame. When blocked on the rover, I look at occupancy age versus feed FPS. --- ### Why Amazon / why switch (honest intern+projects, no fake Amazon worship) **Map.** Common closer; Prince internships+LP; Raghav previous work. Not a 10454435-specific bank. **Spoken.** I am switching from internships and shipped projects into an SDE I role where the bar is production judgment under a clock. Live Code for this loop is Java; I already think in HashMaps, heaps, and graphs of tool nodes — IQVIA’s LangGraph is that shape; Ylogx RLS is authorization on the data path; Horizon is a real-time pipeline with a date. Amazon’s loop (other UTA/AUTA IEs) keeps asking ownership, dive deep, and deliver under constraint. That matches work I already did: three-tier RLS as an intern, mAP gate before 20+ cameras, ERC date with a live costmap. I want to do that at SDE I scope, with Java as the shop language for new services, not because I memorized posters. I will not fake “I have loved Amazon since childhood.” I will not invent an AWS bill I saved. I will not claim I already work like a Bar Raiser. CUSAT CSE 8.42, IEDC Tech Team, eight hackathons including CodeRecet 1st, MLH Best Project, Magnathon 2.0 runner-up — that is the student side. IQVIA Kochi April 2026–present and Ylogx Nov 2024–Oct 2025 are the intern side. Horizon Feb–Jun 2024 is the date side. If they ask why not stay at IQVIA: I am an intern on research systems I am proud of (200+ sites, 200+ page BRDs, LangSmith). SDE I at Amazon is a different scope and a different bar. I will not trash IQVIA or Ylogx. --- ### Current role (GFG 2025 BR) — keep two systems I am a software developer intern at IQVIA in Kochi, April 2026 to present. Two systems: LangGraph Deep Research — Firecrawl, Bing, DuckDuckGo, Google Playwright, rank more than two hundred websites, LangGraph checkpoint; Hybrid RAG — Azure AI Search Hybrid + Semantic, GraphDB, LangGraph adaptive retrieval, LangSmith evals and test-case gen on more than two-hundred-page BRDs. Python FastAPI. Live Code is Java. Before that: Ylogx 99.9% and 40% faster reports; Horizon 17th/80+. If they want three one-liners then one STAR: IQVIA 200+; Ylogx 99.9%; Horizon 17th — then I go deep on one. I do not list StratifyLabs, GiftedBooks, Argus, and eight hackathons in the current-role answer. --- ### Proud of — pick ONE **Map.** Ruchi FTC R2. If intro already used Horizon: GiftedBooks 3–10 min. If not: Horizon 17th. Spare: Argus 73→89. Do not list three. --- ### Completing a project on your own **Map.** LC 7563011 R2. Internships were team. Own-project: Argus, GiftedBooks, or StratifyLabs. Walk problem → architecture → one hard part → resume metric. Do not pick InstaRecon. --- ### Unnamed two-LP block Lead Dive Deep (Ylogx −35% + three tiers) then Deliver Results (Horizon 17th despite stale costmap). If intro used Horizon: Ylogx 40%/99.9% or GiftedBooks 3–10 min. Third if needed: GiftedBooks, one label. Do not guess Job 10454435 titles. --- ## 4 PRACTICE / NOT YOURS **DO NOT SPEAK THESE AS ADARSH.** Fabricated stories get busted. Bar Raisers ask for logs, names, baselines, and the discarded option. These eight skeletons show SHAPE only, at unnamed companies. After each: this is not on your resume. --- ### Skeleton 1 — Missed SLA Situation. At an unnamed company the on-call dashboard promised 99.95 percent monthly availability. Task. Restore the SLA after a region blip. Action. I rolled a config, I wrote a postmortem, I added a monitor. Result. We missed the month and I “learned communication.” This is not on your resume. If you speak it, a BR will ask for logs and you will fail. You do **not** have a missed 99.9% or a missed ERC. Speak Argus 73% as the failed bar, or Horizon sequencing, not this SLA fanfic. --- ### Skeleton 2 — Manager fight Situation. My manager wanted to ship Friday. I wanted tests. Task. Stand my ground. Action. I raised my voice in stand-up, I escalated skip-level, I “won.” Result. We shipped the week after and the manager respected me. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Your conflict is hybrid versus vector-only, or RLS-in-DB versus app-only. No fight scene. No quote. --- ### Skeleton 3 — Mentee of 12 Situation. I led a pod of twelve juniors. Task. Hire and develop. Action. I ran 1:1s, I wrote ratings, I promoted two. Result. Retention went up 20 percent. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. You were IEDC Tech Team and Horizon software. No mentee count. No ratings. Pair on launch files and hackathon modules. --- ### Skeleton 4 — AWS bill saved $40k Situation. Cloud spend was out of control. Task. Frugality. Action. I rightsized instances and saved forty thousand dollars. Result. Finance sent a thank-you. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Your Frugality is Redis −35% bot DB latency and CloudFront/ECS, not a dollar amount. If you do not remember a bill: say so. Never invent one. --- ### Skeleton 5 — Missed ERC / “we almost didn’t show” Situation. Challenge day, the rover wasn’t ready. Task. Hero weekend. Action. I pulled three all-nighters and we barely made it. Result. We missed a run but “learned.” This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Resume: 17th/80+, date met, stale costmap as the setback you *handled*. Do not invent a miss. --- ### Skeleton 6 — Haar cascades then YOLO miracle Situation. We used Haar cascades for PPE, it failed, then I “switched to deep learning.” Task. Are Right. Action. One-line OpenCV then YOLOv8. Result. 95 mAP. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Resume: YOLOv9, 73→89 mAP, 15k images, 24 FPS. No Haar. No 95. No Purplle YOLOv8+ByteTrack as if it were Argus. --- ### Skeleton 7 — SEO 403 as the Dive Deep Situation. www versus non-www returned 403, search console dropped. Task. Dive Deep. Action. I found noindex, I fixed DNS, ranking recovered. Result. Traffic +N%. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Prep-only. On-resume Dive Deep is Ylogx Redis + RLS. Do not lead with 403. --- ### Skeleton 8 — Qdrant gold-set of 10,000 and a named intern drama Situation. My intern rival hid the Qdrant collection, the gold set had ten thousand labeled BRD questions, p95 was 47 ms, cache TTL was 90 seconds. Task. Everything. Action. I confronted them. Result. I became team lead. This is not on your resume. If you speak it, a BR will ask for logs and you will fail. Resume retrieval is Azure AI Search + GraphDB. No Qdrant. No gold-set size. No p95. No TTL. No intern drama. No team-lead title. --- ### Extra shape you will also refuse (same rule) Unnamed company missed customer SLA; “I cut the feature the PM loved”; “I saved the sprint”; student headcount on GiftedBooks; Haar; AWS $; mentee of 12. All of these get the same closer: **This is not on your resume. If you speak it, a BR will ask for logs and you will fail.** --- ## End lock Job **10454435**: still none. R1/R2 done. This file is spoken STAR for later lives. Primary spread: do not break. Metrics: Aug 2026 resume only. Java Live Code; production Python/TS/ROS2. InstaRecon: ethics one-liner, redirect. GiftedBooks GitHub AegisAI: mismatch, resume wins. Ylogx Dive Deep: Redis + RLS, not SEO. Login Tracker: unverified, not in this fragment. Rate Limiter: independent SDE I live count 1, not UTA, not an LP story. If you don’t remember a number: “I don’t remember the exact protocol; the resume number is X; the mechanism was Y.” Never invent a baseline.