# LP + projects — R3 / R4 / HM / BR (Adarsh Vishwakarma) Notes for Adarsh. Not a prediction of Job **10454435**. Other SDE I / UTA / AUTA candidates reported these prompts. Your loop may differ. Unnamed stays unnamed. Still **no** public IE names a live-round LP or project question for this Job ID. **Lock.** R1 (13 Aug 2026) and R2 (18 Aug 2026) are done. This fragment is the **third and fourth live**: Bar Raiser, hiring manager, or a behavioral-heavy technical. Later rounds grill **failures, feedback, missed deadlines, “why that approach,” and 10× scale**. They are not a second recitation of the same success STAR. If they already heard Horizon 17th or Ylogx 99.9% in an intro, go to the failure inside that story, or switch primary per the spread table — do not retell the win. **Java once, then stop.** Live Code is Java. Internships and projects shipped Python (FastAPI, LangGraph, YOLOv9, ROS2), TypeScript (NestJS, React), and ROS2 C++/Python on Horizon. Same data-structure ideas; different syntax. 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. No phishing, credential, or exploit talk. **GiftedBooks GitHub.** The `giftedbooks` README currently describes AegisAI. Mismatch. GiftedBooks answers come from the Aug 2026 resume only. **Metrics lock.** Only numbers on `Resume_Adarsh_Vishwkarma_Aug26.pdf`. No invented miss on ERC or 99.9%. No manager fight scene. No mentee count. No Qdrant. No Haar cascades. Ylogx SEO 403/noindex is prep-only — do not lead Dive Deep with it. --- ## 1. How to STAR in R3/R4 (why-chain, 90s, then stop) Amazon’s later behavioral minutes are not a personality quiz and they are not “tell me the poster again.” The interviewer already has R1/R2 notes. They are checking whether you can **own a failure mode, measure it, and talk about trade-offs without hiding behind “we.”** You get about a minute, sometimes ninety seconds, before they start saying “why.” That is enough for a complete story if you stop listing nouns and start speaking in sentences. Rehearse this skeleton. Do not read the labels in the room. Weave them. ``` Situation (15s): product + who hurt + the constraint that made the obvious path fail. Task (10s): what I owned, in one job-shaped sentence. I, not the team slogan. Action (40–50s): three concrete steps I took, in order, with a mechanism. Result (15s): one resume metric, then stop talking. Then wait. They will ask why. ``` **R3/R4 weight.** Front-load the defect, the disagreement, or the ship-blocker. The result still lands on a resume number, but the body is *what was wrong and what I did*, not a highlight reel. 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. **The why-chain.** Amazon probes with “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. Redis, not a bigger RDS, because the hot path was repeat questions and schema lookups. Hybrid plus graph, not vector-only, because clause IDs are not cosine. Costmap fusion, not another sensor, because the date does not wait on a purchase order. 2. **How did you know it was working.** A resume metric or an eval, not a feeling. Minus 35% bot database latency. 73% → 89% mAP. LangSmith fail-if-no-span. 17th of more than eighty. Do not invent p95, TTL, or a gold-set size the resume does not state. 3. **What would you do differently.** A real missing test or a sequencing miss, not false humility. Denied cross-tier RLS cases from day one. A held-out camera so 89% is not leak. Freeze an eval slice before the fourth search tool. Do **not** invent a missed ERC or a missed 99.9% to sound humble. Stay on one story for the chain. Three whys is a good sign. Reciting sixteen LP names is a bad sign. **I vs we.** Use **I** for the steps you actually took. Use **we** only for a team result you could not have produced alone (Horizon’s 17th place is a team score; Ylogx RLS in the database is an “I designed and shipped” claim). **Then stop.** A 90-second STAR leaves room for their probes. Unnamed LP blocks on other AUTA loops were often 10–20 minutes with two or three prompts; BR/HM can go ~30 minutes on behavioral. If you run long they will cut you. If you run in four-word bullets they will think you memorized a poster. After the metric, shut up. **Do not invent.** Missed ERC, missed 99.9% SLA, manager fight, mentee count, AWS bill, Qdrant, p95, cache TTL, student headcount, Haar cascades. Those are not on the Aug 2026 resume. If you do not have a personal-feedback quote, say so and use Argus 73% mAP as the honest negative signal. If they open a GitHub README that contradicts the resume, the resume wins. **Example why-chain on Dive Deep (stay here if they stay).** Why Redis? Because every natural-language turn was hitting Postgres and the bot was paying for it. Why not a bigger RDS first? Because the hot path was repeat questions and schema lookups, which is a cache, not a larger instance. Why still RLS after the cache? Because a cached answer for the wrong org is a leak, and app-only filters fail on a JOIN. That is three whys on one story. --- ## 2. Primary spread table (do not break) No project is the **primary** story on more than three Leadership Principles. Backups may repeat. If they chain five LPs, change the primary — do not retell GiftedBooks five times. Copy discipline from `interview_bible/11-leadership-principles.html`. Do not promote a backup into a fourth primary. | Leadership Principle | Primary | Backup | Optional | | --- | --- | --- | --- | | Customer Obsession | GiftedBooks | Ylogx SQL RAG +65% | Argus alerts | | Ownership | Ylogx RLS + 99.9% | Horizon ERC software | IQVIA evals | | Invent and Simplify | IQVIA Hybrid RAG + test-case gen | StratifyLabs browser inference | GiftedBooks PYQ | | Are Right, A Lot | Argus 73%→89% | Ylogx measured latency | IQVIA LangSmith | | Learn and Be Curious | IQVIA LangGraph | Horizon ROS2 | 8 hackathons | | Hire and Develop the Best | IEDC CUSAT Tech Team | Horizon teammates | hackathon pair work | | Insist on the Highest Standards | Argus 15k + mAP gate | Ylogx 99.9% / 210ms | GiftedBooks 99.5% | | Think Big | StratifyLabs marketplace + 3D lab | IQVIA 200+ sites | Argus 20+ cameras | | Bias for Action | Horizon ERC date | CodeRecet 1st / 8 events | Ylogx cache ship | | Frugality | Ylogx Redis −35% + CloudFront/ECS | Horizon costmap vs hardware | StratifyLabs browser vs GPU farm | | Earn Trust | GiftedBooks RAG + 99.5% | Ylogx 3-tier RLS | IQVIA LangSmith before ship | | Dive Deep | Ylogx Redis + RLS | IQVIA retrieval | Argus eval | | Have Backbone; Disagree and Commit | IQVIA hybrid vs vector-only | Ylogx RLS-in-DB vs app-only | Horizon software vs extra sensors | | Deliver Results | Horizon 17th/80+ | Ylogx 40% / 99.9% | GiftedBooks 3–10 min | | Strive to be Earth’s Best Employer | Horizon knowledge sharing | IEDC CUSAT | hackathon teammates | | Success and Scale Bring Broad Responsibility | Argus 20+ cameras after 89% mAP | Ylogx ALB/ECS with RLS | IQVIA evals so wrong answers do not scale | **Counts (do not break).** Ylogx is already at three primaries (Ownership, Frugality, Dive Deep). IQVIA is at three (Invent, Learn, Backbone). Argus is at three (Are Right, Highest Standards, Success and Scale). Horizon is at three (Bias for Action, Deliver Results, Earth’s Best Employer). GiftedBooks is at two (Customer Obsession, Earn Trust). StratifyLabs is Think Big primary. IEDC CUSAT is Hire and Develop primary. Do not promote a backup into a fourth primary. Customer-product trust (grounded RAG, isolation) is Customer Obsession. How honest you are about your own work is Earn Trust. “Leave things better” is Success and Scale, not Ownership. --- ## 3. Default unnamed pair if they do not name the LP Other SDE I / UTA / AUTA write-ups often log “two LPs, unnamed.” Job **10454435** still has none. Do not guess their titles in the room. Lead with this pair. If they want a third, GiftedBooks doubt time (hours to 3–10 minutes) is Customer Obsession / Earn Trust backup — pick **one** label and go deep, do not list all three. If they already used Horizon in the intro, **do not** retell 17th as the second unnamed LP. Switch Deliver Results to Ylogx 40% / 99.9% or GiftedBooks 3–10 min, and keep Horizon as the metric they already have. Depth on the stale costmap is fine if they ask *why* the rover placed; a second victory lap is not. ### Dive Deep — Ylogx Redis + 3-tier RLS, −35% bot DB latency (not SEO) 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. GFG 2025 named “Dive Deep.” Many unnamed rounds are “last time you deep-dived a bug.” Same story. Do **not** lead with Ylogx SEO 403/noindex; that is prep-only. **Spoken (~90s).** At Ylogx I was a software intern from November 2024 to October 2025 on a full-stack AI business-intelligence product: FastAPI and NestJS over Postgres, React dashboards, and a LangChain SQL RAG chatbot that turned natural language into SQL. On a demo the chatbot looked finished. It returned rows. That was the anecdote: “it works.” Users who were not SQL experts could ask a question in English. The metrics disagreed. Every natural-language turn was hitting the database, so the bot path was slow in a way that felt like “the model is thinking” when the database was the cost. And if isolation lived only in the application, a JOIN or a generated query could outrun a UI filter and return another organization’s rows. My job was not to add another prompt. My job was to find whether slowness was query shape, missing cache, or security checks done in the wrong layer, and to make correctness and isolation real for three organizational tiers on the same data path the bot and the reports already shared. Measuring that path myself was not beneath the intern ticket. I measured bot database latency instead of celebrating that SQL came back. 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 a generated query 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. Then I put Redis on the hot bot path so repeat questions and schema lookups were not a round trip every time. The cache did not skip RLS. A cached answer still had to belong to the right tier. Bot database latency dropped 35 percent. Data-analysis productivity for people who were not writing SQL went up 65 percent. Report generation got 40 percent faster. We held 99.9 percent uptime, and the dashboards stayed in the sub-210 millisecond number on the resume. Thirty KPI dashboards later showed a 60 percent ops-efficiency move. The deep dive was latency plus isolation, not a search-console ticket. **Tiny example.** A user in one org asked a dashboard-style question that, without RLS, would have joined a shared fact table and returned another org’s KPIs. The generated SQL was syntactically fine. The database role with RLS is what made it fail closed. Redis then stopped us from paying for that same allowed query on every follow-up turn. **R3/R4 probes (stay on this story).** Why Redis not a bigger RDS? Repeat NL and schema lookups are cache-shaped. Why still RLS after the cache? A cached leak is still a leak. What failed first? Assuming the generator was wrong when retrieval from Postgres and isolation in the wrong layer were the bugs. Differently: treat RLS policies as tested artifacts — denied cross-tier cases — from day one, the same way I later treated Argus mAP as a gate. I will not invent a p95 or a TTL; the resume number I will defend is minus 35 percent. If they push SEO 403/www versus non-www, that is prep-only and not this story. ### Deliver Results — Horizon 17th/80+ (backup if intro already used the rover) Official Deliver Results: key inputs, right quality, timely fashion; despite setbacks, rise to the occasion and never settle. Default second unnamed LP. Other loops named Deliver Results, strict deadline, or “work under pressure.” Same Horizon story. The setback is stale costmap / Gazebo ≠ dirt, **not** an invented missed ERC. **Spoken (~90s) — use this if they have not already heard the rover.** From February to June 2024 I was on the core software team for Team Horizon at CUSAT, building a semi-autonomous Mars rover for the European Rover Challenge 2024. More than eighty teams were competing. The event date does not move because your Gazebo world is still ugly. 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. A lab video was not a result. A rover that maps in bags and still collides is not a result either. The setback was time-sync: the operator feed could look live while occupancy was already old. I shipped a real-time camera pipeline at 60 FPS with GStreamer instead of a naive web stream that would have dropped the operator feed. The supporting 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 sensor fusion in front of the actuators so we were not waiting on extra sensors as the default spend. 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 so two million points per second did not all try to leave the rover. We placed 17th globally out of more than eighty teams. Obstacle detection accuracy improved 40 percent on the resume’s pipeline work. Collision risk dropped 55 percent after costmap fusion, not after buying another sensor. 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. **If intro already used Horizon — backup spoken (~90s), Ylogx 40%/99.9%.** At Ylogx the result that mattered was not “the chatbot demoed.” It was reports 40 percent faster at 99.9 percent uptime, with non-technical users actually able to ask questions. I owned the serving and data path: FastAPI plus NestJS plus Postgres, Redis on the bot path for minus 35 percent database latency, CloudFront and ECS so dashboards stayed in sub-210 milliseconds without putting an LLM on that hot path. Isolation was part of the result: three-tier RLS, because a fast leak is not a delivered result. Analysis productivity plus 65 percent and thirty KPI dashboards with plus 60 percent ops efficiency sit on the same product. If they want a customer-facing backup instead, GiftedBooks took student doubts from hours to three to ten minutes at sub-300 milliseconds and 99.5 percent uptime. **R3/R4 probes.** Perception without planning was the first failure class — a fast video that still collides. I would time-box “one extra sensor versus fusion” against a collision-risk number earlier. I would sequence 60 FPS and costmap on a calendar to the ERC date even earlier. Do not invent a cut on 99.9 percent. --- ## 4. R3 / HM / BR exact prompts from bible §5.D — full spoken answers Every prompt below is **IE-asked** in Question-Research-BIBLE.md §5.D (plus the Fleet extras that section points at). They are **not** from the July 2021 Leadership Principles PDF. That PDF is sixteen definitions. Job **10454435** still names none of these. Unnamed stays unnamed. When the prompt is vague (“conflict,” “bug,” “deadline”), still answer with the official LP **behavior** — Dive Deep is metrics vs anecdote, Ownership is not “that’s not my job,” Backbone is disagree then commit wholly. These are speakable answers, not a script. Rehearse out loud. After the metric, stop. ### Current role — IQVIA two systems **Source.** GFG 2025 Bar Raiser. Prince-style internships-plus-LP rounds want three one-liners then one STAR. Arijit R1 wanted academic or internship unnamed. **Spoken.** I am a software developer intern at IQVIA in Kochi, from April 2026 to the present. I work on two research systems, not on “we called an LLM.” The first is a LangGraph Deep Research platform: a stateful agent graph that fans out Firecrawl plus Bing, DuckDuckGo, and Google Playwright, ranks sources across more than two hundred websites, and writes a result. The second is Hybrid RAG: Azure AI Search with hybrid and semantic retrieval, a GraphDB, and LangGraph choosing retrieval mode on more than two-hundred-page BRDs and PDFs, with LangSmith tracing for evals and generated test cases. Production is Python FastAPI. Live Code for this loop is Java. Before IQVIA I interned at Ylogx from November 2024 to October 2025 on full-stack AI BI: FastAPI, NestJS, Postgres, SQL RAG, three-tier RLS, Redis, and AWS CloudFront/ECS, with 99.9 percent uptime and 40 percent faster reports. Before that I was core software on Team Horizon for ERC 2024, 17th of more than eighty teams. If they want one number after the IQVIA minute, I give 200+ sites and 200+ page BRDs and stop. If they want Prince’s three one-liners: IQVIA 200+; Ylogx 99.9 percent; Horizon 17th — then I go deep on one. I do not list StratifyLabs, GiftedBooks, Argus, and eight hackathons in the current-role answer. Those wait for “tell me about a project.” **R3/R4.** A current-role dump that names six products fails first. Keep two systems and two numbers. If they ask Java at IQVIA: I did not ship this as a JVM service; I would wrap the same graph of nodes in Java services if that were the shop. If they already heard IQVIA in R1/R2, lead with the ranking bug or the hybrid-versus-vector disagreement, not the architecture slide again. ### Last time you deep-dived a bug **Source.** GFG 2025 BR. Same story as official Dive Deep. Not SEO. **Spoken.** The last deep dive I will defend on the resume is Ylogx, not a search-console 403. The chatbot returned SQL and rows, so it looked healthy in a demo. That was the anecdote. The metrics disagreed. Every natural-language turn was hitting Postgres, so latency felt like “the model is thinking” when the database was the cost. And if isolation lived only in the application, a JOIN or a generated query could outrun a UI filter and return another organization’s rows. My task was to find whether slowness was query shape, missing cache, or security in the wrong layer, and to make three organizational tiers real on the path the bot and the reports already shared. Measuring that path was not beneath the intern ticket. I measured bot database latency instead of blaming the generator. I put RLS and RBAC in Postgres so generated SQL ran as the user’s database role. I put Redis on the hot path for a 35 percent drop in database latency without skipping RLS. A cached answer still had to be the right tier. Uptime stayed 99.9 percent. Analysis productivity for non-SQL users moved 65 percent because the bot was both faster and safe to trust. Report generation on the same product got 40 percent faster. Dashboards stayed in the sub-210 millisecond number. I will not lead this answer with www versus non-www indexing. That is prep-only. **Tiny example.** A JOIN that is legal SQL and illegal for the tenant is the bug class. The SQL can look perfect in my account and still be a leak. The database role is the fix. Cache is the follow-up so we do not pay for the allowed query every turn. Showing the SQL to the user is how wrong grain shows up even when the tenant is correct. **If they already heard Ylogx.** IQVIA ranking contradiction: two high-ranked pages disagree; first-hit retrieval looks confident; rank by agreement and source type; refuse a span-less sentence; LangSmith showed empty scrapes still feeding the writer. Or Argus 73 percent mAP as an eval bug. Or Horizon stale costmap: operator at 60 FPS, planner on last second’s rocks. SEO 403 is still prep-only. ### Conflict with a coworker or manager **Source.** GFG 2025 BR; GFG April 2026 R3/HM. Technical. No invented fight. **Spoken — IQVIA hybrid vs vector-only.** 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 technical, because that is what I actually had. At IQVIA the fast path for “RAG on PDFs” is a vector index and a demo. A teammate or stakeholder who wants that path is not being foolish. They are optimizing for a ship date. I still had to say the failure mode out loud. Two-hundred-page BRDs have clause IDs and requirement graphs that cosine search misses. Test-case generation from the wrong section is expensive for whoever consumes the BRD. Similar embeddings are not coverage. I argued for Hybrid plus Semantic Azure AI Search plus a GraphDB plus LangGraph adaptive retrieval, with LangSmith traces so we could see a bad path. Lexical search catches “R-141.” Semantic search catches “what is the retention policy.” Graph hops catch dependencies. Once we chose that path, I committed by instrumenting evals and test-case generation instead of reopening the tool debate every week. Disagree on retrieval. Commit on evals. There is no named interpersonal incident on the resume. **If they already heard IQVIA — Ylogx RLS-in-DB vs app-only.** The conflict is row-level security in Postgres versus app-only filters. I would not ship the bot on app-only checks because a JOIN leaks. The commit was three-tier RLS and RBAC as a platform rule, not a UI hide. Optional Horizon: software costmap versus “buy another sensor.” We committed to fusion. Collision risk dropped 55 percent. **Tiny example.** “Ship the vector demo by Friday” versus “R-141 must be retrievable on Monday.” The disagreement is coverage, not personality. The commit is an eval that fails if a generated test case has no overlapping span with the retrieved section. I did not overrule a named manager. I will not perform a fight. **R3/R4.** They may push “what did the other person say.” Stay on the failure mode and the eval table. I would put the eval table on the table before the argument so commit is on numbers. ### Last negative feedback **Source.** GFG 2025 BR; Rudraksh BR; LC 6653463 R3. No invented quote. **Spoken.** The honest on-resume version is Argus, and I will say that out loud if they want a person’s name I do not have. 73 percent mAP was the feedback. It was not production quality for PPE on a floor. At that number you miss helmets and you also fire junk boxes. Either failure trains people to ignore the alarm. I did not treat 73 percent as a starting blog metric. I did not ship that detector as the standard for alerts. I trained and evaluated on more than 15,000 images, iterated YOLOv9 to 89 percent mAP, and required 24 FPS so the model we called “right” could keep up with the stream. Then I logged events to Postgres and containerized a path that could support more than twenty cameras. Safety violations later dropped 50 percent and compliance doubled, after the gate, not before it. I do not have a confirmed manager quote, and I will not invent one. If they insist on person-feedback and I have no review story, I say so and stay on this metric. Backup if they already heard Argus: the system’s own feedback on Ylogx was bot database latency before Redis, which we cut 35 percent. **Tiny example.** A missed helmet in odd lighting is a silent fail. A false box is a loud fail. Both live inside 73 percent mAP. A one-line color threshold does not take PPE from 73 to 89 at 24 FPS. An LLM describing the frame is not the feedback response either. The response is data, a held-out eval, and a frame-rate gate. I will not invent a train/test split the resume does not name. ### Learned something not required / not part of the job **Source.** GFG 2025 BR; LC 6653463 R2. Same as Learn and Be Curious. Rudraksh “quickly learn something new” is the same primary, shorter clock. **Spoken.** LangGraph, LangSmith, and Azure hybrid search were not required to “call an LLM from FastAPI.” That call was already a skill from Ylogx SQL RAG. They were required to rank more than two hundred websites and to retrieve from two-hundred-page BRDs without inventing citations. Tutorial-complete is not architecture. I learned enough to draw a planner, parallel tools, a ranker, a writer, and a fail-if-no-span eval. I learned stateful graphs so a failed scrape does not restart the whole run. 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. The result is the IQVIA platform 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. If they want a different stack, Horizon ROS2, GStreamer at 60 FPS, and ZED 2 at more than two million points per second were not the web intern job either. That learning showed up as 17th of more than eighty teams, plus 40 percent obstacle detection, minus 55 percent collision risk. **Tiny example.** I discarded “one vector index named RAG.” Long BRDs needed lexical plus semantic plus graph. I also discarded a linear chain for Deep Research, because retries and parallel tools are a graph. Those discards are the learning, not a course certificate. **R3/R4.** Tutorial-complete without an eval failed first. I would freeze a gold set of questions before adding a third search tool. What I read: LangGraph state and Azure hybrid docs until I could draw the graph, then traces — not a certificate. ### Found an issue in a product even though it wasn’t your task **Source.** IE.in 2024-grad AUTA R2. Exact wording. LC 8362604 “problem I identified myself” is the same spine. **Spoken.** On paper my Ylogx ticket was a chatbot and a report builder. Isolation and bot database latency were not “just SQL generation.” They were still the product, because 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 the database so generated SQL could not outrun the app. I measured bot database latency instead of assuming the model was slow. I shipped Redis for a 35 percent drop without skipping RLS. 99.9 percent uptime held. Analysis productivity moved 65 percent. That is an issue that was not the demo ticket and still had to close. If they already heard Ylogx, GiftedBooks is the backup. Doubt hours were the product issue even if the named work was VR labs and avatars. Students were the customers. I shipped contextual RAG, sub-300 milliseconds, 99.5 percent uptime, and PYQ topics. Doubts went to three to ten minutes. **Tiny example.** The SQL looked right in my user account. A JOIN across orgs would not have looked right in production. Finding that while building the bot is the “not my task” part: platform isolation was not the chatbot demo. The UI filter was the tempting ticket-sized fix. It was the wrong layer. ### Urgent requirement / trade-off / strict deadline / bias for action **Sources.** IE.in 2024-grad AUTA R2 (urgent/trade-off). LC 7724048 R1 and R2 (strict deadline). GFG sde-1-17 (tough deadline plus compromise). LC 6570344 (bias for action). Dhananjai R2 (struggled to meet a deadline; why it mattered). Same Horizon spine. Do not invent a miss. **Spoken — one story, four prompt wordings.** Horizon’s ERC date was the urgent requirement and the strict deadline. More than eighty teams show up whether or not your Gazebo world is pretty. Extra hardware would have been the comfortable trade: wait for another sensor, hope the field gets easier. I traded for software on the sensors we had. That is also bias for action: 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. I shipped a 60 FPS GStreamer feed so we were not blind. I put costmap planning and fusion on the ZED 2 at more than two million points per second. The compromise was software planning versus more hardware, **not** “skip safety.” I refused to trade away a live costmap. A fast camera with no planning is not a trade. It is a miss that still looks like motion. Collision risk dropped 55 percent. Obstacle detection improved 40 percent. We placed 17th of more than eighty. The field did not match Gazebo lighting. Frames that were fine in sim arrived late in dirt. The urgent work was pipeline latency versus a stale occupancy grid, in the weeks we had, not a purchase order. Downsampling for viz and keeping density for the local costmap was the same-week trade. Why the deadline mattered: the challenge does not accept a rain check. A new architecture in the last week would have been motion. A stable feed plus costmap was the deadline. I did not cut isolation on a later job to make a date, and I will not invent a cut on Ylogx 99.9 percent. Reckless action would have been skipping the costmap to look busy. I did neither. **If they already heard the rover.** IQVIA: hybrid retrieval versus ship-vector-only under document pressure. What I cut was extra Playwright renders and extra tools. What I did not cut was citations and traces. Ylogx intern version of bias for action: measure bot database latency, ship Redis, minus 35 percent, do not wait for a rewrite. Eight hackathons, first at CodeRecet, MLH Best Project, Magnathon 2.0 runner-up, is the shortest clock. Resume wording is rapid software development and production-ready deployment. I will not invent hours-to-prize. ### Outside your scope / designated work / responsibility outside your own work **Sources.** LC 7850431; LC 6570344; GFG sde-1-17. Same family. **Spoken.** On Horizon I was software team, not “the camera person” on an org chart that did not exist. GStreamer and mapping were still ship-blockers for ERC 2024, so I owned them. A webcam-quality demo stream was the designated work for a lab video. Sixty frames per second on the operator and autonomy path was outside that demo and inside the challenge. ZED 2 mapping at more than two million points per second into a costmap was the same expansion. Outcome: the operator could see, the planner had occupancy, we placed 17th of more than eighty, obstacle detection up 40 percent, collision risk down 55 percent. If they already heard the rover, Ylogx RLS was outside a narrow chatbot ticket. Three-tier isolation as platform, 99.9 percent held, minus 35 percent bot latency, plus 65 percent analysis productivity. If they want IQVIA: test-case generation from BRDs expanded RAG into something a human could audit, with LangSmith traces, on more than two-hundred-page documents. **Tiny example.** A Flask MJPEG stream was in scope for a screenshot. GStreamer at 60 FPS was outside that screenshot. The outcome is not a prettier README. The outcome is a feed the challenge can use. Public `Gstreamer-UDP` is a supporting webcam artifact in the same family, not a second project I will pad this answer with. **R3/R4.** Staying inside a ticket while the rover is blind failed first. I would name the ship-blocker in the first week, out loud, so “out of scope” is a schedule fact, not a surprise. ### What’s a task you’re most proud of? — pick ONE **Source.** Ruchi FTC R2. In the room, pick **one**. Do not list Argus as a third unless they already used Horizon and GiftedBooks. **Default pick for R3/R4 if intro already used Horizon: GiftedBooks.** If they have not heard the rover, Horizon is the cleaner Deliver Results proud-of. **Spoken — GiftedBooks (customer).** The task I am proud of is GiftedBooks RAG question-and-answer that took student doubts from hours to three to ten minutes. I am proud of it because the customer is a student with their own PDF, not an internal dashboard. VR labs without answers are decoration. A generic bot that sounds like a textbook is the wrong customer. Students were losing hours even with the notes in front of them. I grounded Q&A in the upload so answers came from their material. I held sub-300 millisecond API responses and 99.5 percent uptime so the assistant was usable during study, not only in a demo. I added PYQ topic suggestions so people studied what exams actually ask, which is analytics, not an LLM guessing the syllabus. Reading efficiency went up 35 percent. Engagement went up 50 percent. Comprehension moved 2.5 times. GitHub currently mismatches AegisAI on that repo. I am proud of the resume product, not the wrong README. I would pick this when they want Customer Obsession or Earn Trust under “proud.” **Spoken — Horizon (date + team).** The task I am proud of is the Horizon software stack that actually ran at ERC 2024. I am proud of it because the date was real, the field was not Gazebo, and the score is 17th of more than eighty teams. I was core software from February to June 2024 on a semi-autonomous Mars rover. The hard part was not a model card. It was a 60 FPS GStreamer feed plus ZED 2 mapping at more than two million points per second plus a costmap the planner could trust when lighting and dust disagreed with simulation. Stale occupancy was the bug class: the rover plans on yesterday’s world and looks autonomous while it collides. I treated that as a ship-blocker. We fused sensors and planned on a live grid. Obstacle detection improved 40 percent. Collision risk dropped 55 percent. I would pick this when they want Deliver Results or Bias for Action sitting under “proud.” **R3/R4.** Listing three proud projects fails first. I pick one, give the metric, and stop. Spare, not a third recitation: Argus 73 to 89 percent mAP. ### Describe a mistake you made. What did you learn? **Source.** Rudraksh BR. Do not invent a missed SLA. **Spoken.** The mistake I will own is treating 73 percent mAP on Argus as progress. It looked like a training curve going the right way. It was not a standard I should alert humans on. At 73 percent you miss helmets and you fire junk boxes. Either one trains people to ignore PPE alarms. I almost treated a notebook number as a ship bar. I learned to put an mAP gate and a 24 FPS gate in front of scale. I went to more than 15,000 images and 89 percent mAP before more than twenty cameras. I logged alerts to Postgres so compliance was an audit trail. Violations down 50 percent and compliance doubled are what happened after I stopped calling 73 percent a win. I will not invent a missed 99.9 percent or a missed ERC to sound more humble. Those misses are not on the resume. Backup mistake if they already heard Argus: SQL RAG at Ylogx without a cache. The chatbot “worked.” The database paid for every turn. I learned to measure bot database latency and ship Redis for minus 35 percent, still behind RLS. **Tiny example.** Shipping 73 percent to one webcam is a notebook. Shipping it to twenty cameras is alarm fatigue. The mistake was the scale decision, not the first training run. An LLM looking at frames would have been a second mistake on the same product. ### Significant technical challenge **Source.** GFG April 2026 R3/HM. Raghav R1 “most difficult project” is the same pick-one rule. **Spoken.** IQVIA Hybrid RAG on more than two-hundred-page BRDs was the significant technical challenge, and I will go deep on this one unless they already heard it. The constraint was confidential long documents plus a need to rank and cite. Stuffing the PDF into a context window blows tokens and invents citations. Vector-only retrieval returns similar paragraphs and misses clause IDs. A linear chain cannot retry a scrape or hop a requirement graph. Test cases generated from the wrong section waste whoever consumes the BRD. I built hybrid plus semantic Azure AI Search, GraphDB hops, LangGraph adaptive retrieval, and LangSmith traces. Lexical for IDs, semantic for policy language, graph for dependencies, cite or abstain. More than two hundred sites on the Deep Research side, more than two-hundred-page documents on the RAG side. No extra percentage on the resume. I did not use Qdrant. The resume is Azure AI Search and GraphDB. If they want physics instead of retrieval, Horizon is the challenge: ZED 2 at more than two million points per second plus a costmap that must not go stale, 60 FPS, 17th of more than eighty. If they want eval, Argus is 73 to 89 percent mAP at 24 FPS on more than 15,000 images. I pick one and go deep. I do not list all three. **Tiny example.** Two pages disagree; the writer still sounds sure. Ranking by agreement and refusing a span-less sentence is the challenge, not GPU count. Empty scrape still feeding the writer is the LangSmith-visible version of the same bug. ### Limited information **Source.** GFG April 2026 R3/HM. **Spoken.** Deep Research at IQVIA has no single gold page. You rank more than two hundred websites without knowing which one is right. Limited information is the job, not an excuse. I used parallel scrape and search — Firecrawl, Bing, DuckDuckGo, Playwright — because one source family fails on different sites. I ranked on quality, recency, and agreement instead of first-hit confidence. I used LangSmith so a fluent paragraph without a source is a fail. I refused to answer without a retrieved span. Two high-ranked pages that contradict each other is the limited-information case in one picture: you do not get to pretend you know. No extra percentage. I will not invent a labeled gold-set size the resume does not state. Backup: Horizon field sensors versus lab. Gazebo did not tell us the dust or the lighting. We still had to plan. Fusion on ZED 2 plus a live costmap, collision risk down 55 percent, 17th of more than eighty, was how we acted without a perfect world model. 60 FPS was how the operator saw the same incomplete world in time. **R3/R4.** Trusting the first snippet failed first. I would add source-type weights on the ranker earlier. ### Customers + teammates — two products **Source.** Bhavya R2. Two different projects. Do not mash them. **Spoken — customers (GiftedBooks).** I will give two stories on two products, because that prompt asks both. Customers first. On GiftedBooks, students were losing hours on doubts even with VR labs and PDFs in front of them. A pretty lab that cannot answer from their material is decoration. I shipped contextual RAG over the upload, held sub-300 milliseconds and 99.5 percent uptime, and added PYQ topics so people studied what exams ask. Doubts went to three to ten minutes. Reading efficiency went up 35 percent. Engagement went up 50 percent. Comprehension moved 2.5 times. That is above and beyond a VR lab that only looks like school. If they already heard GiftedBooks, Ylogx SQL RAG plus 65 percent analysis productivity is the employer-customer version: non-technical users stopped waiting on an analyst for every slice, behind three-tier RLS. **Spoken — teammates (IEDC / Horizon).** Teammates second. I was on the Tech Team at IEDC CUSAT and on Horizon software. I am not a manager and I do not have a junior headcount or ratings. I shared environment setup, reviews, and “here is the path we already burned.” On Horizon I made GStreamer, ZED, and costmap knowledge usable by the rest of the rover team, because a solo notebook does not place 17th of more than eighty. At eight hackathons, including first at CodeRecet, I paired on the risky module instead of hoarding it. CGPA 8.42 is education, not this answer’s metric. **Tiny example.** Customer story ends on three to ten minutes. Teammate story ends on someone else launching the GStreamer node. If I use one metric for both, they will think I have one story. I will not mash GiftedBooks into IEDC. ### Aditya trio — Ownership / pressure / learn (three projects) **Source.** Aditya R2: Ownership; working under pressure; learning and adapting to new technologies. Three short spoken answers. Three projects. Do not reuse Ylogx three times. **Ownership — Ylogx.** I owned three-tier RLS and RBAC in Postgres on the path the bot and reports share. I was an intern, not a manager, and app-only filters were still not ownership. Generated SQL ran as the user’s database role so a JOIN could not outrun the UI. I kept 99.9 percent uptime while report generation went 40 percent faster. I put Redis on the bot path for minus 35 percent database latency so isolation was not an excuse for load. Analysis productivity moved 65 percent. Thirty dashboards and plus 60 percent ops efficiency sit on the same serving path, sub-210 milliseconds. I still owned the data path. **Working under pressure — Horizon.** ERC 2024 does not slip. More than eighty teams show up. I shipped 60 FPS and a live costmap on the hardware we had instead of waiting on extra sensors. Stale occupancy was the pressure bug: the rover looks autonomous while it collides. We fused ZED 2 at more than two million points per second and planned on a live grid. 17th of more than eighty, collision risk down 55 percent, obstacle detection up 40 percent. Pressure without a costmap is just stress. Backup pressure: eight hackathons, first at CodeRecet, production-ready deploy in the window. **Learning and adapting — IQVIA.** LangGraph, hybrid Azure search, GraphDB, and LangSmith were new relative to Ylogx FastAPI and NestJS. I adapted by making traces the stop condition on more than two hundred sites and two-hundred-page BRDs. I learned adaptive retrieval instead of one vector index. I did not adapt by adding tools without evals. Test-case generation is itself a traced job. Production is Python. Live Code is Java. **If they rush.** Ylogx RLS and 99.9 percent; Horizon 17th of more than eighty; IQVIA LangGraph on 200+ page BRDs. Then they pick one to probe. I do not volunteer sixteen LP names. Three LPs on one project fails Aditya’s structure. ### Complex problem that required POCs and multiple solutions **Source.** LC 7563011 HM AUTA. LC 7981646 Innovate/Frugality architecture slot uses the same drawing. **Spoken.** IQVIA Deep Research and Hybrid RAG were the POC series, and I will walk what I tried and what I discarded. One scraper is not enough. Firecrawl, Bing, DuckDuckGo, and Playwright each fail on different sites, so the graph fans out and checkpoints state. Vector-only is not enough on two-hundred-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, so that POC died. Stuffing the PDF into one prompt dies on tokens and invented citations. Evals in LangSmith were the stop condition: if a generated test case has no overlapping span, the path is not done. I discarded Qdrant as a story because the resume is Azure AI Search, not a prep synonym. The result is more than two hundred sites ranked and more than two-hundred-page BRDs processed with traces and test-case generation. No extra percentage. Backup: Ylogx was FastAPI plus NestJS plus Postgres plus Redis plus RLS versus “just LangChain.” The discarded path was a service-role LLM writing SQL. App-only filters were another discarded POC. Three-tier RLS, minus 35 percent bot latency, 99.9 percent uptime is what shipped. **Tiny example.** Playwright on every URL is a POC that loses on cost. Bing snippets first, full render only when the snippet is thin — that is a discarded-versus-kept pair I can draw. Invent: adaptive retrieval plus GraphDB. Frugal: cache scrapes and fewer renders. I will not train a foundation model. ### Missing a deadline — honest lock **Sources.** LC 8362604 BR; LC 7563011 R2; LC 7724048 BR; LC 6475219. Resume does **not** state a missed ERC or a missed 99.9 percent SLA. Do not invent a miss. **Spoken.** 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 to hear how you flag risk, not a fictional failure. What I will say is how the immovable date was met and how I would communicate if a ship-blocker were late. ERC 2024 was fixed. More than eighty teams show up. I sequenced 60 FPS and the costmap so the run happened. We placed 17th of more than eighty. Collision risk dropped 55 percent. Obstacle detection improved 40 percent. 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. If I later confirm a real miss off-resume, I will not use it until it is on the record. **Tiny example.** “We might miss ERC” is useless in week twelve. “Costmap is planning on stale occupancy; fusion is the path; extra sensor is the cut” is communication. The run happened because 60 FPS and planning were sequenced first, not because I pretend we missed. **R3/R4.** Hiding a stale costmap until the last week would have been the miss. Sequencing 60 FPS and planning first is how I avoid inventing a failure the resume does not have. 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. ### Why that approach / what would you do differently / how did you measure **Sources.** Rushikesh BR; igreaper BR (multi-level why, impact, differently, disagreements, measure success). This is a **probe pattern**, not a new project. Answer on whichever STAR they are already in. Do not restart. **Spoken pattern.** I answer on the story they are in. Discarded option plus metric. Then stop. - **IQVIA.** I chose hybrid plus graph because vector-only misses IDs and dependencies. I would freeze an eval slice earlier. Success is 200+ page BRDs processed with traces and test-case gen, not fluency. Disagreement is Backbone: hybrid versus vector-only, then commit on LangSmith. - **Ylogx.** I chose RLS in the database because app-only fails on JOINs. I chose Redis because repeat NL is cache-shaped, not a bigger RDS. I would test denied cross-tier cases from day one. I measured success as minus 35 percent bot latency, 99.9 percent uptime, and three tiers, not a feeling. I will not invent p95 or TTL. - **Argus.** I chose not to scale 73 percent mAP. I would hold out a camera earlier. Success is 89 percent mAP, 24 FPS, minus 50 percent violations, compliance 2×. - **Horizon.** I chose costmap fusion over extra hardware. I would time-box that choice against collision risk earlier. Success is 17th of more than eighty and minus 55 percent collision risk. - **GiftedBooks.** I chose RAG plus small generate over stuffing the PDF. I would add a citation eval (not claimed on the resume — I would add it). Success is hours to 3–10 minutes, sub-300 ms, 99.5 percent uptime. **Tiny example.** “Why Redis not bigger RDS?” Repeat NL questions are cache-shaped. Success is minus 35 percent, not an instance class I will not invent a price for. If they are already asking that, I do not restart a new STAR unless they change LP. ### Didn’t know what to do next / comfort zone / stuck **Sources.** LC 7850431 BR; LC 7280347; LC 7406809; GFG intern “completely stuck.” Same as Learn, with the stuck beat first. **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, and SQL RAG. A tutorial was not a graph. I broke it into research agents, a ranker, hybrid retrieve, then a LangSmith stop condition. Parallel tools and retries are a graph, so a linear chain died. Vector-only died on clause IDs. The stop condition is what made “I don’t know” finite: fail if no overlapping span, including generated test cases. Result: 200+ sites ranked, 200+ page BRDs processed. Horizon ROS2 backup if they want a non-LLM stuck: GStreamer clocks versus ZED rate versus costmap, 17th of more than eighty. ### BR bundle — challenge / identified myself / simplifying a process **Source.** LC 8362604 AUTA 2026 BR (also: missing a deadline — use the honest block above). Rate-limit/geofencing were **their** story follow-ups, not your Rate Limiter LLD. **Challenge I solved.** Argus mAP gate, or IQVIA hybrid on 200+ page BRDs. Pick one. See significant-challenge and mistake cards. **Problem I identified myself.** Ylogx latency/RLS while the chatbot “worked,” or Argus 73 percent as a fail. See “not your task.” **Simplifying a process.** IQVIA test-case generation from BRDs so humans are not handwriting cases from 200+ pages. StratifyLabs browser inference, ML iteration −30 percent, versus a GPU loop per tweak. GiftedBooks PYQ instead of full-PDF grind (+35 percent reading efficiency). Redis versus a bigger database (−35 percent). Simplifying without a mechanism is a deleted feature; these have resume numbers. ### Feedback / critical issues / went beyond your task **Source.** LC 6806195 AUTA BR. Same three as 7850431 BR. **Feedback.** Argus 73 percent mAP. **Critical.** Ylogx isolation/uptime (wrong data is a critical issue) or Horizon collision risk. **Beyond.** Horizon GStreamer or IQVIA evals. Speak them as three short paragraphs, three projects, then stop. ### Completing a project on your own **Source.** 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 LP blocks (do not guess titles) | Prompt | Source | What to do | | --- | --- | --- | | two LP UNNAMED (went poorly) | GFG 2025 R2 | Default pair. Depth on Action, not titles. | | two LP UNNAMED deep follow-ups | Naina UTA R2 | Default pair + one backup. Follow-ups on why/metric. | | eight LP UNNAMED | Vanshika Interview 2 | Start default pair, then Customer Obsession + Ownership. Do not recite 16 names. | | LP first ~15 min UNNAMED | Akash AUTA R1 | Default pair. Short STAR. | | 15–20 min generic LP then SD | Arijit AUTA R2 | Default pair; be ready to stop LP. Do not eat the hour. | | intense ~30 min behavioral UNNAMED | Karun University TA | §5 depth; they may only code after. Metrics ready. | | all 16 LP STAR ~15–20 min each UNNAMED | Nisarg AUTA BR | One primary each from the spread. No project reused >3. They will probe 2–3. | | 2 LP UNNAMED (several AUTA/OA-as-R1) | Naina, 6800853, Rajprem, etc. | Default pair. | | internships + LP UNNAMED | Prince | IQVIA + Ylogx + Horizon one-liners then one STAR. | | BR LP-only / formality UNNAMED | Reddit AUTA Chennai; LC 7406809 | Default pair, still use numbers. Recruiter “3 LP only” still got DSA on LC 7406809 — be ready to stop talking. | | LP 5 UNNAMED / BR 3–4 deep dives | LC 8014509 R2; LC 8029194 | Spread table. Change primary. | | collaboration / ownership UNNAMED | Uday BR | Horizon team + Ylogx Ownership. | | project critical work | LC 6475219 R3 | Argus or Horizon. | | Innovate / Frugality with architecture | LC 7981646 extra | Invent IQVIA + Frugality Ylogx Redis/ECS. | **OA-as-R1 trap.** If an IE numbered OA as “Round 1,” their “Round 3” is often this loop’s **R2 analogue**, not Adarsh’s upcoming R3. Flag it in your head; still use these answers if the prompt matches. ### GenAI should / should-not (short) **Sources.** LC 7850431 R3; LC 7724048 R3/HM; LC 8014509; Vaishali; Deepak KFC-orders. Same slot may also be DSA (Distance K, k-th largest, Next Greater, LCA). If they write a heap or a tree, write the algorithm. Full GenAI cards live in the 01 fragment / Answer-BIBLE §6. Eight to twelve sentences here. **Spoken (~10 sentences).** I use GenAI where the output is high-volume, pattern-based, and cheaply verifiable. At IQVIA that is Deep Research across more than two hundred ranked sites and Hybrid RAG on more than two-hundred-page BRDs, with LangSmith traces, cited spans, and generated test cases a human can spot-check. At Ylogx that is SQL RAG: natural language to SQL behind three-tier RLS, which is plus 65 percent analysis productivity. At GiftedBooks that is PDF Q&A grounded in the upload, sub-300 milliseconds, 99.5 percent uptime. I verify in three ways: overlapping span with retrieved chunks or the writer abstains; SQL executed as the user’s RLS role, not as a service superuser; gold questions with expected citations or expected SQL shape. I keep dashboards off the model: Ylogx sub-210 milliseconds is not an LLM path, and Redis minus 35 percent is efficiency on the bot path. I do not dump ninety days of rows into a prompt — I pre-aggregate in SQL, same shape as “highest KFC orders last three months.” I do not use GenAI for Ylogx RLS policy, for Horizon costmap or actuation, or for Argus PPE boxes: those must be exact and auditable. An LLM that describes a helmet is not 89 percent mAP at 24 FPS, and Copilot for Live Code is draft-then-compile and bounds-check, not paste. Fluency is not correctness. If this slot is actually Distance K labeled Fluency, I write the tree algorithm and stop the essay. --- ## 5. All 16 LPs — 60s primary STAR each (resume-true), backup named Speak in paragraphs. Do not dump this table in the room. After the metric, stop. R3/R4: if they already heard the win, start on the defect inside the same primary. ### Customer Obsession — primary GiftedBooks (2/2 GiftedBooks with Earn Trust) Amazon means: start with the customer and work backwards; earn and keep customer trust; obsess over customers, not competitors. **Spoken (~60s).** GiftedBooks is a VR learning suite: 3D labs, AI avatars, and a study path for students who already have PDFs and still lose hours on doubts. A pretty lab that cannot answer a question from the student’s own material is decoration. A generic chatbot that sounds like a textbook is also the wrong customer. The student is trying to pass an exam with *their* notes, not with a model’s prior. I needed to cut time-to-answer without inventing content. I shipped RAG over uploaded PDFs so question-and-answer was contextual. Sub-300 millisecond API responses and 99.5 percent uptime were product constraints, not vanity. I added prioritized topic suggestions from previous-year-question analysis so students studied what exams actually ask. Reading efficiency improved 35 percent. Engagement improved 50 percent. Content comprehension moved 2.5 times faster. Average doubt resolution went from hours to three to ten minutes. GitHub currently shows an AegisAI README. I will not mix that product into this story. **Backup.** Ylogx SQL RAG chatbot: non-technical users, plus 65 percent data-analysis productivity. **Optional.** Argus alerts so floor staff see PPE violations; safety violations −50 percent, compliance 2×. **R3/R4.** PDF Q&A that is slow or generic failed first. First pass without PYQ still left people guessing *what* to study. I would put a small RAG eval set — wrong-citation versus grounded — in earlier. The resume does not claim GiftedBooks had LangSmith; I would add it. Do not invent student headcount. ### Ownership — primary Ylogx (1/3) Amazon means: owners; long-term value over short-term results; never “that’s not my job.” “Leave things better” is Success and Scale, not this LP. **Spoken (~60s).** Ylogx was a full-stack AI BI intern role. The chatbot and the report builder would have been wrong *and* a leak if organizational boundaries were only an application check. I was not the named security team. I was the intern shipping the data path. That is exactly when people say “filter in the UI and ship.” I owned correctness and isolation for three org tiers, not “chatbot works on my user.” I put RLS and RBAC for three organizational tiers in the data path the bot and the reports share. Generated SQL ran as the user’s database role. I kept the BI app at 99.9 percent uptime while report generation went 40 percent faster. I added Redis on the bot path for a 35 percent drop in database latency so security did not mean “add a check and ignore load.” Forty percent faster reports, 99.9 percent uptime, plus 65 percent analysis productivity, minus 35 percent bot database latency, thirty KPI dashboards, plus 60 percent ops efficiency, sub-210 milliseconds. **Backup.** Horizon: GStreamer 60 FPS and the costmap as ship-blockers; 17th of more than eighty. **Optional.** IQVIA evals and tracing, not “the agent returned text.” **R3/R4.** App-only role checks fail closed in a demo and open in a JOIN. I would treat RLS policies as tested artifacts from day one. If they ask whether I was a manager: no. Ownership here is the data path. ### Invent and Simplify — primary IQVIA (1/3) Amazon means: invent and simplify; not limited by “not invented here”; accept being misunderstood. **Spoken (~60s).** I am a software developer intern at IQVIA in Kochi, April 2026 to present. 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. 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. 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 for adaptive retrieval. Lexical search catches clause IDs. Semantic search catches “what is the policy.” Graph hops catch “what depends on this requirement.” LangSmith tracing, evals, and test-case generation made the invention checkable. A citation-gated writer is the simplify: fewer fluent sentences, more exact coverage. I did not use Qdrant. More than two hundred websites ranked. More than two-hundred-page BRDs processed. No extra percentage. **Backup.** StratifyLabs browser inference; ML iteration −30 percent; marketplace 50+. **Optional.** GiftedBooks PYQ; +35 percent reading efficiency. **R3/R4.** Vector-only on long BRDs retrieves similar paragraphs, not the requirement graph. I would freeze an eval slice of BRD sections to expected test cases before adding another search tool. ### Are Right, A Lot — primary Argus (1/3) Amazon means: right a lot; seek diverse perspectives; disconfirm beliefs. “Right” is an eval you would alert a human on. **Spoken (~60s).** 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 notebook looked fine. The metric said it was not right on a floor. I treated 73 percent mAP as a fail, not a starting blog metric. I trained and evaluated on more than 15,000 images and iterated YOLOv9 until 89 percent mAP. I ran the detector at 24 FPS so the “right” model was also the one that kept up with the stream. I wrote events to Postgres so compliance was an audit trail, not a GIF. I did not replace boxes with an LLM looking at frames. Seventy-three to 89 percent mAP, 24 FPS, safety violations down 50 percent, more than twenty camera feeds, compliance doubled. **Backup.** Ylogx measured latency: 40 percent reports, sub-210 ms, Redis −35 percent. **Optional.** IQVIA LangSmith before treating research as done. **R3/R4.** 73 percent mAP was the first fail. I would lock a held-out 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. ### Learn and Be Curious — primary IQVIA (2/3) Amazon means: never done learning; curiosity needs a stop condition or it is wandering. **Spoken (~60s).** Ylogx had been FastAPI, NestJS, Postgres, and SQL RAG. 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. I used LangGraph for stateful multi-agent Deep Research so a failed scrape did not restart the whole run. I used hybrid plus semantic search plus GraphDB because a long BRD is not a single embedding query. I used LangSmith tracing and evals so curiosity had a pass/fail, including generated test cases. I discarded the idea that “RAG” means one vector index. I discarded a linear chain for Deep Research because retries and parallel tools are a graph. The platform as on the resume: more than two hundred sites ranked, more than two-hundred-page BRDs, evals and test-case generation. **Backup.** Horizon ROS2 / GStreamer / ZED; 17th of more than eighty, +40 percent obstacle, −55 percent collision. **Optional.** Eight hackathons; first at CodeRecet, MLH Best Project, Magnathon 2.0 runner-up. **R3/R4.** Treating RAG as one vector index failed first. I would write a one-page “what we will not retrieve” list earlier so agents do not wander two hundred sites without a ranker. ### Hire and Develop the Best — primary IEDC CUSAT Tech Team (1/3 on IEDC/education) Amazon means: raise the bar; develop leaders; coaching. Honesty: Software Team Member, Team Horizon; Tech Team, IEDC CUSAT. **Not** a people manager. No invented reports, ratings, or headcount. **Spoken (~60s).** 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. I was a team member, not a manager. I did not hire anyone. I do not have a mentee count. 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, because a solo notebook does not place 17th. 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 is education, not this LP’s metric. **Backup.** Horizon teammates: unblock camera and mapping. **Optional.** Same eight-hackathon set, named as collaboration. **R3/R4.** Siloed knowledge was the first failure. I would write a ten-line runbook earlier. The resume does not claim a wiki, so I will not say we had one. I will not invent ratings or an Amazon-style hiring bar I did not run. ### Insist on the Highest Standards — primary Argus (2/3) Amazon means: relentlessly high standards; defects do not get sent down the line; problems stay fixed. **Spoken (~60s).** 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. I used more than 15,000 images, not a toy set. I moved YOLOv9 from 73 to 89 percent mAP. 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. Safety violations down 50 percent. Compliance doubled. The standard was the gate, not a paragraph in a README. **Backup.** Ylogx 99.9 percent uptime, sub-210 milliseconds, CI/CD on ECS and Docker. **Optional.** GiftedBooks 99.5 percent uptime and sub-300 milliseconds. **R3/R4.** 73 percent mAP failed first. On Ylogx, a chatbot that hits Postgres on every natural-language turn failed first. I would publish fail cases — missed helmet, leaked tier, timeout — as the standard, not only the happy metric. ### Think Big — primary StratifyLabs (1/3) Amazon means: thinking small is a self-fulfilling prophecy; look around corners for customers. On an intern resume this is “make the default path reusable,” not a fake TAM. **Spoken (~60s).** Computer-vision work gets stuck in a local loop: one GPU, one notebook, one model, one person. 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. I built a 3D simulation lab for real-time experimentation and prototyping. I put inference in the browser so ML iteration dropped 30 percent. 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. Iteration time down 30 percent. More than fifty marketplace items. I will not add a go-to-market number. **Backup.** IQVIA Deep Research across more than two hundred ranked websites. **Optional.** Argus on more than twenty cameras — Argus is already at three primaries, so keep this optional. **R3/R4.** Local-only inference made every experiment a machine problem. I would draw a clearer “marketplace versus your fine-tune” boundary earlier. Do not invent users. ### Bias for Action — primary Horizon (1/3) Amazon means: speed matters; calculated risk; an irreversible collision is not the kind of decision you rush. **Spoken (~60s).** Team Horizon, February to June 2024. ERC 2024’s date is fixed. We were building a semi-autonomous Mars rover on ROS2 with student hardware. 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. I shipped the 60 FPS GStreamer camera feed instead of a slow MJPEG experiment. I got ZED 2 mapping at more than two million points per second into RViz and Gazebo. I put costmap planning and predictive fusion on the sensors we had. Field time was part of the action: Gazebo lighting was not dirt. 17th globally of more than eighty teams. Obstacle detection up 40 percent. Collision risk down 55 percent. **Backup.** Eight hackathons; first at CodeRecet; MLH Best Project; Magnathon 2.0 runner-up. **Optional.** Ylogx: measure bot database latency, ship Redis, minus 35 percent. **R3/R4.** A map without a costmap failed first. I would time-box “one sensor versus fusion” with the collision-risk number earlier. Speed without the 60 FPS feed would have been the wrong bias — acting on a blind rover. ### Frugality — primary Ylogx (2/3) Amazon means: accomplish more with less; no extra points for budget size. Not cheapness. Not spending on the wrong layer first. **Spoken (~60s).** 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. I put Redis on the bot path and cut database latency 35 percent. 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. Cached answers still had to be tier-correct under RLS. I will not invent an AWS bill or a TTL. Minus 35 percent bot database latency, sub-210 milliseconds, 99.9 percent uptime, 40 percent faster reports. **Backup.** Horizon: costmap and fusion, collision risk down 55 percent, versus adding hardware. **Optional.** StratifyLabs browser inference, iteration down 30 percent, versus everyone needing a training box. **R3/R4.** Uncached SQL RAG failed first. I would write cache invalidation rules down with RLS. The resume does not detail TTL; I will not invent one. A dashboard KPI that must be sub-210 milliseconds does not go through an LLM at all. ### Earn Trust — primary GiftedBooks (2/2) Amazon means: 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 you are about your own work. Keep GiftedBooks as primary so Ylogx does not take a fourth slot. **Spoken (~60s).** GiftedBooks is a VR learning suite: 3D labs, AI avatars, and PDFs. The awkward sentence I had to say out loud is that a pretty lab that cannot answer a question from the student’s own material is decoration, and a generic chatbot that sounds like a textbook is not a teaching assistant. That is vocally self-critical. I will not pretend the VR surface was enough. I said the failure mode before we dressed it up: hallucinated citations and downtime both burn trust. I shipped contextual PDF Q&A, not a general LLM with a textbook voice. I held sub-300 millisecond API responses and 99.5 percent uptime. I used PYQ-prioritized topics so the product is honest about exams. Per-user PDFs must not retrieve another student’s notes. Doubt time from hours to three to ten minutes. Reading efficiency up 35 percent, engagement up 50 percent, comprehension 2.5 times. I will not claim zero incidents. I will not say the body of work smelled of perfume because the VR looked expensive. **Backup.** Ylogx three-tier RLS: speak candidly that a JOIN leaks if isolation is only in the UI. **Optional.** IQVIA LangSmith before research goes out. Argus 73 percent is Are Right / Highest Standards primary; do not steal it as a fourth Argus primary here. **R3/R4.** Pretending the VR lab was enough failed first. I would add the same citation-eval habit as IQVIA on GiftedBooks. Not claimed on the resume — I would add it. If they want Customer Obsession, same product, different angle: start with the student and work backwards. ### Dive Deep — primary Ylogx (3/3 Ylogx primaries full) Amazon means: all levels; details; audit; skeptical when metrics and anecdote differ; no task is beneath you. Lead with Redis and RLS, not SEO. **Spoken (~60s).** 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. The metrics disagreed: every natural-language turn was hitting Postgres, and a JOIN could outrun an app filter. I measured bot database latency instead of celebrating that SQL came back. I put row-level security and RBAC in Postgres for those three tiers so generated SQL ran as the user’s database role. 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. Bot database latency dropped 35 percent. Analysis productivity went up 65 percent. Report generation got 40 percent faster. 99.9 percent uptime held. The deep dive was latency plus isolation. It was not a search-console ticket, and I will not lead with 403 or noindex. **Backup.** IQVIA: wrong chunks on two-hundred-page BRDs; hybrid plus semantic versus vector-only plus GraphDB plus LangSmith. **Optional.** Argus 73 percent mAP — dive into data and eval, not a bigger backbone first. **R3/R4.** Assuming the generator was wrong failed first. I would want denied RLS probes as tests. The resume does not name p95; I will say latency was measured enough to claim minus 35 percent. ### Have Backbone; Disagree and Commit — primary IQVIA (3/3 IQVIA primaries full) Amazon means: respectfully challenge; do not compromise for social cohesion; once a decision is determined, commit wholly. No invented manager blow-up. **Spoken (~60s).** 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. I stated the failure mode: similar embeddings are not requirement coverage. I proposed hybrid plus semantic plus GraphDB plus LangGraph adaptive retrieval. I committed via LangSmith evals and test-case generation. Once we chose the path, I traced it rather than reopening the tool debate every week. Commit wholly means instrument the chosen path, not keep a quiet side project on FAISS. Hybrid RAG as on the resume. Two-hundred-page BRDs. Evals and test-case generation. Disagree on retrieval; commit on evals. There is no named interpersonal incident on the resume, and I will not invent one. **Backup.** Ylogx: RLS in the database versus app-only filters. Commit: three-tier RLS and RBAC as the platform rule. **Optional.** Horizon software costmap versus “buy another sensor”; collision risk down 55 percent. **R3/R4.** Vector-only failed first. I would write the decision as an eval table before the argument so commit is on numbers. I did not overrule a named manager. ### Deliver Results — primary Horizon (2/3) Amazon means: key inputs, right quality, timely fashion; despite setbacks, never settle. Output, not activity. **Spoken (~60s).** ERC 2024 had more than eighty international teams. I was Horizon software from February to June 2024. 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. 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. 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 resume does not state a miss, and I will not invent one to sound humble. **Backup.** Ylogx: 40 percent faster reports, 99.9 percent uptime, sub-210 milliseconds. Use this if intro already used the rover. **Optional.** GiftedBooks hours to 3–10 minutes. **R3/R4.** Perception without planning failed first. I would sequence demo-risk items on a calendar to the ERC date even earlier. Do not invent a cut on 99.9 percent. ### Strive to be Earth’s Best Employer — primary Horizon (3/3 Horizon primaries full) Amazon means: safer, more productive environment; are teammates growing, empowered, ready for what’s next. Honesty: teammate environment on a student rover team, not Amazon HR. No retention survey. Do not collapse this into the same sentence as Hire and Develop — here the bar is growth and ready-for-what’s-next, not “I coached.” **Spoken (~60s).** Horizon rover software is unusable if camera and mapping knowledge lives in one head. ERC is a team score. I was a software team member, not a people manager, not HR, and not a hiring bar raiser. 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. I worked as a software team member, not a hero module. I shared GStreamer, ZED, and costmap constraints so others could integrate. I used the same habit at IEDC and at hackathons: the person next to you can run the build. 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. I did not run a mentoring program with a name. **Backup.** IEDC CUSAT Tech Team. CGPA 8.42 is not this LP’s metric. **Optional.** MLH, CodeRecet, and Magnathon teammates; no junior headcount. **R3/R4.** Siloed ROS nodes failed first. I would pair earlier on launch files. If they want Amazon-scale DEI or parental leave: I will not fake it. This LP on my resume is teammate environment. ### Success and Scale Bring Broad Responsibility — primary Argus (3/3 Argus primaries full) Amazon means: 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. **Spoken (~60s).** 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. I did not scale 73 percent mAP. I got to 89 percent mAP at 24 FPS on more than 15,000 images. Then I containerized more than twenty feeds, wrote Postgres logs, and shipped alerts that cut violations 50 percent and doubled compliance. Video stays off the database; events go in. More than twenty cameras, minus 50 percent violations, 2 times compliance, after the eval gate, not before it. I will not invent EKS because Ylogx is the ECS story. I will not claim a climate or community program. **Backup.** Ylogx ALB plus ECS plus CloudFront: scale the BI path that already has RLS. 99.9 percent, 210 milliseconds. Scaling SQL RAG without cache and RLS scales leaks. **Optional.** IQVIA evals so ranked research and generated tests do not scale garbage across two hundred pages and sites. **R3/R4.** Scaling a 73 percent model failed first. I would add a per-camera error budget. The resume does not state it; I would add it. --- ## 6. Project 2-minute pitches A project question is not a tour of every intern and GitHub repo. Problem → architecture → one hard part → resume metric → stop. Then they ask why. Four likely why-questions sit under each pitch. **10× users:** honest (cache, RLS, evals, mAP gate) — not fake Kafka HLD. If it is not on the Aug 2026 resume, it is not in the pitch. **Java vs production (say once).** Live Code is Java. These products did not ship as JVM services. ### IQVIA Deep Research **Problem.** Analysts cannot research the open web by hand for every request. One scraper and first-hit “confidence” ship a contradiction. A one-shot prompt is not coverage across 200+ sites. **Architecture.** FastAPI plus a LangGraph: planner → parallel tools (Firecrawl, Bing, DuckDuckGo, Google Playwright) → ranker (quality, recency, **agreement**) → synthesizer. State on the graph so a failed scrape retries instead of restarting the run. LangSmith traces every node. **One hard part.** Ranking, not a named outage. Two high-ranked pages contradict. First-hit retrieval looks confident. The writer will narrate a fluent compromise that neither page said. I ranked by agreement and source type and refused to answer without a retrieved span. LangSmith showed wrong tool order and empty scrapes still feeding the writer. No extra percentage. No Qdrant. **Resume metric.** 200+ websites ranked. Stateful orchestration. Evals. No invented latency SLA. **Four likely why-questions.** 1. **Why four tools, not one scraper?** Each family fails on different sites. Parallel plus checkpoint is cheaper than restarting the run. 2. **Why agreement ranking, not first-hit?** Two pages disagree; first-hit is false certainty. 3. **Why LangGraph, not a linear chain?** Retries and parallel tools are a graph. 4. **What would you do differently?** Freeze gold questions with expected citations before adding Playwright. Citation-required writer in the type system: no span, no sentence. **10× users / 10× research jobs.** Honest: cache scrapes; do not re-crawl the same URL every run; fewer Playwright renders, more Bing snippets when cost spikes; tighter recency; the ranker and the eval set have to scale or you multiply fluent garbage. Not Kafka, not a private embedder I trained, not an invented QPS. ### IQVIA Hybrid RAG **Problem.** 200+ page BRDs and PDFs do not fit in a context window. Stuffing them invents citations. Hand-written test cases from a two-hundred-page requirements document do not scale. Vector-only retrieval returns similar paragraphs and misses clause IDs. **Architecture.** Chunk by structure (clause IDs, tables, headings). Azure AI Search Hybrid + Semantic: lexical for “R-141,” semantic for “what is the retention policy.” GraphDB hops for “what depends on this requirement.” LangGraph adaptive retrieval chooses the mode. Writer is citation-gated. Test-case generation is itself a traced LangSmith job. Resume retrieval is Azure AI Search + GraphDB, not Qdrant. **One hard part.** Similar embeddings ≠ requirement coverage. Test cases generated from the wrong section waste whoever consumes the BRD. The Backbone argument was hybrid plus graph plus traces versus ship-vector-only. Commit was evals, not another library debate. **Resume metric.** 200+ page BRD/PDF processed; stateful agents; evals; test-case generation. No extra percent. **Four likely why-questions.** 1. **Why not dump the PDF into the prompt?** Tokens, invented citations, no eval. 2. **Why GraphDB, not chunks only?** Dependencies are a graph, not cosine. 3. **Why hybrid, not FAISS-only?** Exact IDs and table names miss on vector-only. 4. **How do you know it works?** Gold questions with expected citations. Fail if answer has no overlapping span. Humans spot-check a sample of generated tests. **10× documents / 10× questions.** Honest: 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; evals so wrong answers do not scale. 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. JWT/RLS theatre belongs on Ylogx. ### Ylogx ConvBI / SQL-RAG **Label.** Resume product is the full-stack AI BI intern work (Nov 2024–Oct 2025): custom report builder, LangChain SQL RAG, 3-tier RLS, Redis, 30 dashboards, CloudFront/ECS. Conv-BI on GitHub is **Ylogx-adjacent** (Power-BI-like report builder), not a separate job and not a second resume bullet. Warpflow is a GitHub UI experiment, not the shipped orchestrator. Do not lead with SEO 403/noindex. **Problem.** Operators need custom reports, a natural-language chatbot over warehouse facts, and live KPI tiles — without one tenant ever reading another. A free-form LLM over raw rows is a token and correctness failure. App-only `WHERE org_id = ?` dies the first time a JOIN forgets the filter. **Architecture.** NestJS: auth, RBAC, report CRUD. FastAPI: SQL RAG (NL → constrained SQL). Postgres is source of truth; generated SQL runs **as the user’s DB role** so RLS applies. Redis cache-aside on schema and repeat answers; keys are tier-correct. React/Recharts, 30 KPIs. CloudFront → ALB → ECS (Docker), CI/CD. GoDaddy → Route 53 → ALB. REST, not GraphQL as the first cut. LLM never holds a superuser connection string. **One hard part.** The chatbot “worked” while bot DB latency and cross-tier risk were the product. I measured latency, put RLS+RBAC on three tiers, and put Redis in front. Second class: SQL that is syntactically fine but wrong grain — run as the RLS role and show the SQL. Do not lead Dive Deep with www vs non-www 403. **Resume metrics.** 40% faster reports; 99.9% uptime; +65% analysis productivity; −35% bot DB latency; 30 dashboards +60% ops efficiency; sub-210 ms. **Four likely why-questions.** 1. **Why RLS in the database, not app-only?** A JOIN outruns a UI filter. Defense in depth. 2. **Why Redis, not a bigger RDS?** Repeat NL and schema lookups are cache-shaped. −35%. 3. **Why Postgres, not Mongo?** Joins, aggregations, transactions, RLS. 4. **Why is the dashboard not an LLM path?** Sub-210 ms is cache plus pooling. Frugality is also “do not put a foundation model on the hot path.” **10× users / 10× NL questions.** Honest: 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, k8s is on the skills list but I was not cluster owner. The production path I will defend is ECS + Docker + CloudFront. LangSmith belongs to IQVIA. Login Tracker is unverified. I did not ship a Rate Limiter (independent SDE I live count = 1, not UTA). ### Team Horizon (ERC 2024) **Problem.** A semi-autonomous Mars rover had to run on challenge day against 80+ teams. A lab video is not a result. A rover that maps in bags and still collides is not a result. Gazebo lighting is not dirt. The date does not move. **Architecture.** ROS2 graph: GStreamer camera at 60 FPS (UDP-family; public `Gstreamer-UDP` is a webcam analogue, not a second bullet); ZED 2 at 2M+ pts/s into RViz/Gazebo; costmap path planning plus sensor fusion; downsample for viz, dense cloud for the local costmap. Perception, mapping, and planning are separate nodes. mAP belongs to Argus. StratifyLabs is a browser CV lab — do not merge stacks. **One hard part.** Stale costmap. Operator sees 60 FPS; planner sees last second’s rocks. Debug is GStreamer pipeline latency versus ROS2 callback versus ZED rate — three clocks. Extra hardware was not the first fix. Software fusion on the ZED 2 we had was. Resume does not state a missed ERC. **Resume metrics.** 17th globally / 80+ teams; 60 FPS; 2M+ pts/s; +40% obstacle detection; −55% collision risk. **Four likely why-questions.** 1. **Why ROS2, not a giant vision model?** ERC is localization + planning + comms. A weights file is one node. 2. **Why GStreamer, not TCP MJPEG?** 60 FPS needed a pipeline, not a Flask stream. 3. **Why costmap fusion, not another sensor?** Date and student hardware. −55% is software on the sensors we had. 4. **What failed in the field?** Sim lighting ≠ dirt. Time-sync between feed and occupancy. **10× “users.”** 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. ### Argus **Problem.** Real-time PPE compliance and attendance on a plant floor. 73% mAP is a model card, not a standard you page a supervisor on. Wrong alerts at 20+ cameras scale alarm fatigue. Color/HOG OpenCV will not take PPE across that gap. An LLM describing a frame is not mAP. **Architecture.** Cameras → ingest workers → YOLOv9 (+ NMS) → alert service → Postgres events/attendance. OpenCV for decode/resize, not the detector. Containerized path for 20+ feeds. Video stays out of the database. GitHub `argus-stream-api-server` README is 404 — no invented routes. Purplle-style YOLOv8 + ByteTrack is prep-only; resume is YOLOv9 15k / 73→89. **One hard part.** Treating 73% mAP as progress. A heavier model looks better on a laptop and dies at 20 streams. Tune until 24 FPS held. Lighting and vest color are dataset work (15,000+ images), not a one-line threshold. I will not invent a train/test split. **Resume metrics.** 73% → 89% mAP; 15,000+ images; 24 FPS; violations −50%; 20+ cameras; compliance 2×. **Four likely why-questions.** 1. **Why not ship 73%?** Misses and junk boxes train people to ignore alarms. 2. **Why YOLO, not a heavier two-stage detector?** 24 FPS on 20+ cameras is the constraint. 3. **Why Postgres, not video files?** Audit trail of what fired. Events, not blobs. 4. **Why not an LLM on frames?** Fluency is not mAP. Boxes must be auditable. **10× cameras.** Honest: do not scale 73% mAP; workers per stream or batch; do not run 20 copies of the fattest model if 24 FPS breaks; mAP gate then multiply; per-camera error budget is would-add, not claimed. Not EKS (Ylogx is ECS). Not 95 mAP. Not precision/recall beyond mAP. ### GiftedBooks **Problem.** Students lose hours on doubts even with VR labs and PDFs in front of them. A pretty lab that cannot answer from *their* material is decoration. Dumping the PDF into long context misses the latency budget and invents citations. GitHub currently describes AegisAI — resume only. **Architecture.** VR client → API (auth) → PDF ingest → chunk/embed → retrieve + small generate (sub-300 ms) → PYQ topic ranker **separate from the LLM**. Per-user PDFs must not retrieve another student’s notes. Embed once per PDF; do not re-embed on every question. **One hard part.** RAG latency versus quality. Chunks too large blow sub-300 ms. Chunks too small miss the table on page forty. Uptime 99.5% is hosting, not a model trick. I will not invent a named outage or a student headcount. **Resume metrics.** Sub-300 ms API; 99.5% uptime; reading +35%; engagement +50%; comprehension 2.5×; doubts hours → 3–10 min. **Four likely why-questions.** 1. **Why RAG, not whole-PDF context?** Tokens, latency, invented citations. 2. **Why PYQ as analytics, not an LLM syllabus guess?** Honesty about what exams ask. 3. **Why 99.5% and sub-300 ms as product constraints?** A study assistant that times out during exams is not a product. 4. **What would you add?** A small wrong-citation vs grounded eval (IQVIA habit). Not claimed — I would add it. **10× students / 10× PDFs.** Honest: embed once; cache retrieval; keep generate small; per-user isolation so you do not retrieve another student’s notes. Do not invent MAU. Do not steal LangSmith onto this bullet. Do not mix AegisAI. ### StratifyLabs **Problem.** CV work stuck in a local loop: one GPU, one notebook, one person. Iteration *is* the product. GitHub README is default Next.js — resume wins: stratifylabs.design. **Architecture.** Next.js + WebGL 3D sim lab; browser-based inference; marketplace 50+ pretrained models/datasets; community profiles; URDF editor; Gemini RAG voice bots as 3D characters conditioned on the scene. Small models in-browser; big train jobs off the request path. Deep RLS stays on Ylogx. No LangSmith on this bullet. **One hard part.** Browser inference vs native: WebGL/WASM memory, model size, “works on my GPU box.” Honest class of bug. Do not invent a Jira ticket or a crash percent. Horizon Gazebo is a different product — do not merge. **Resume metrics.** −30% ML iteration time; 50+ marketplace models/datasets. **Four likely why-questions.** 1. **Why browser, not Colab-per-person?** Iteration time. −30% is time, not a cluster. 2. **Why WebGL URDF, not native Gazebo-only?** CV people who are not ROS still prototype. 3. **Why a marketplace?** Reuse versus a private notebook. 50+ is a catalog, not 50 GPUs per user. 4. **What would you add?** Eval harness for Gemini bots grounded in **current** sim state. Clearer marketplace vs personal fine-tune. No GTM number. **10× users.** Honest: 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 cluster size or Stratify RLS. --- ### Off-resume (confirm) Do **not** lead with these unless you confirm they are true and you want them on the record. Prep mapped them; they are **not** Aug 2026 resume bullets. | Story | Why off-resume | If confirmed, which LP | | --- | --- | --- | | Django CUSAT student portal security audit (unauthenticated media) | Prep only | Ownership / Dive Deep / out of scope | | Ylogx SEO 403 / noindex www vs non-www | Prep only | Dive Deep — **do not** use as primary; on-resume Dive Deep is Redis+RLS | | Purplle-style CCTV YOLOv8 + ByteTrack + OSNet Re-ID | Prep only; Argus on resume is YOLOv9 15k / 73→89 | Learn / Are Right — use **Argus** unless you confirm Purplle | | LLM uncensored / abliteration / LoRA / DPO / mergekit | Prep only | Learn — weak for Amazon; skip unless asked about safety | | Hospital JWT + rotating refresh | Prep only | Ownership / Earn Trust | | SuperTokens + MSG91/2Factor + DLT/TRAI OTP | Prep only | Customer Obsession / Earn Trust | | Dubai Mall navigator (SVG + Dijkstra) | Prep only | Customer Obsession | | ConvBI / Warpflow as named **separate** products | Not resume bullets. Conv-BI GitHub exists — label **GitHub / Ylogx-adjacent** | Prefer **Ylogx SQL RAG +65%** | | Qdrant on the BRD system | Prep; resume is Azure AI Search + GraphDB | Do not say Qdrant | | InstaRecon internals | Ethics one-liner only | Not an LP story | | GiftedBooks GitHub AegisAI README | Mismatch | Never mix into GiftedBooks STAR | | StratifyLabs default Next.js README extras | Not resume | Do not invent | | Haar cascades as a “we tried first” | Not on resume | Do not say | --- ## 7. Room checklist - **Round.** R1/R2 done. This is R3/R4/HM/BR. Failure, feedback, deadline, why, 10× — not a second victory lap. - **Unnamed pair.** Dive Deep = anecdote vs metric on Ylogx (minus 35 percent plus three tiers). Deliver Results = Horizon 17th of more than eighty despite stale costmap. If intro already used Horizon: Ylogx 40%/99.9% or GiftedBooks 3–10 min. Third: GiftedBooks 3–10 minutes (Customer Obsession, or Earn Trust if they want self-critical). Pick one label. - **Spread.** Primary ≤ 3. Ylogx / IQVIA / Argus / Horizon are full. Do not promote a backup. - **Aditya trio.** Ownership Ylogx / pressure Horizon / learn IQVIA. Three projects. - **Proud.** Pick Horizon **or** GiftedBooks, not both, unless they ask for two. - **Conflict.** Hybrid versus vector-only, or RLS-in-DB versus app-only. No fight scene. No manager quote. - **Mistake / negative feedback.** Argus 73 percent mAP. No invented quote. If they insist on a person and you have none, say so and stay on the metric. - **Missed deadline.** Do not invent one. Sequence 60 FPS plus costmap. Flag a ship-blocker early and specific. Cut extra hardware/features, not isolation or a 73% safety model. - **Why-chain.** Three whys on the same story, then stop. Discarded option + resume metric. - **GenAI.** IQVIA, Ylogx SQL RAG, GiftedBooks PDF Q&A — yes, with checks. RLS, costmap, Argus boxes — no. If the slot is Distance K, write the tree. - **10×.** Cache, RLS, evals, mAP gate. Not fake Kafka HLD. Not invented QPS. - **InstaRecon.** Ethics one-liner, redirect. - **GiftedBooks GitHub.** Resume only. AegisAI is a mismatch. - **Ylogx Dive Deep.** Redis + RLS. Not SEO 403/noindex. - **Java once.** Live Code Java; production Python/TS/ROS2. - **Hire and Develop / Earth’s Best Employer.** IEDC + Horizon teammates. No mentee count. No fake people-management. - **I, then metric, then shut up.** 90 seconds. Let them ask why. Official LP text: `Amazon+Leadership+Principles.pdf` (1 July 2021). Job **10454435**: still none. Login Tracker: unverified. Rate Limiter SDE I live count: 1 (not UTA). OA is not a round. Unnamed stays unnamed.