12. Genuine STAR — sixteen LPs and Bar Raiser probes

1. Why that approach instead of the obvious one.
2. How did you know it was working (resume metric or eval, not a feeling).
3. What would you do differently (missing test or sequencing, not a fake miss).

Why Redis, not a bigger RDS? Every natural-language turn was hitting Postgres. Repeat questions and schema lookups are cache-shaped. A larger instance would have paid for the same round trip more expensively. I will not invent a bill or a TTL. The number I defend is minus 35 percent bot database latency.

Why still RLS after the cache? A cached answer for the wrong org is a leak. App-only WHERE org_id = ? fails the moment generated SQL writes a JOIN the UI never showed. Generated SQL ran as the user’s database role. Redis keys had to be tier-scoped. Fail-open cache never means fail-open isolation.

What would I do differently? Treat denied cross-tier cases as tests from day one, the same way I later treated Argus mAP as a gate. I assumed the generator was wrong first. Retrieval and isolation in the wrong layer were the bugs. I will not lead this with SEO 403. That is prep-only.

Why not ship 73 percent? At that number you miss helmets and you fire junk boxes. Either failure trains people to ignore PPE alarms. Instinct said the notebook looked fine. The metric said it was not right on a floor.

Why YOLO one-stage, not a two-stage detector? Twenty-four FPS across more than twenty cameras. A heavier detector that dies on the stream is not “more right.” Fluency describing a frame is not mAP.

How did I know? Seventy-three to 89 percent mAP on more than 15,000 images, 24 FPS, then scale. Violations down 50 percent, compliance doubled, after the gate. I do not have a manager quote. I will not invent one. Differently: lock a held-out camera earlier so 89 percent is not leak. The resume does not name that split; I would add it, I am not claiming I already published it.

Why not vector-only? Two-hundred-page BRDs have clause IDs. Cosine returns similar “retention” paragraphs and misses R-141. Requirement dependencies are a graph hop, not a neighbor.

Why GraphDB plus Azure hybrid, not another embedder? Lexical for IDs and table names. Semantic for policy language. Graph for “what else breaks.” I did not use Qdrant. The resume is Azure AI Search and GraphDB.

Commit? LangSmith: a generated test case with no overlapping span is a fail. Once we chose the path I instrumented it. I did not keep a quiet FAISS side project. Differently: freeze an eval slice of BRD sections to expected tests before adding a fourth search tool. No extra percentage on the resume.

Why not buy another sensor? The ERC date does not wait on a purchase order. Extra hardware as the first move is delay dressed as prudence. Collision risk down 55 percent came from fusion on the ZED 2 we had, not from a new unit.

Why GStreamer 60 FPS first? A naive TCP MJPEG stream drops the operator. If the feed slips, the operator is blind. If the costmap is stale, the planner is lying. Those two were sequenced because they were ship-blockers.

How did I know? 17th of more than eighty teams. Obstacle detection up 40 percent. Collision risk down 55 percent. Differently: time-box “one extra sensor versus fusion” against the collision number even earlier. I will not invent a missed ERC.

Situation. GiftedBooks is a VR learning suite: 3D labs, AI avatars, and a study path. Students already have PDFs and still lose hours on doubts. A pretty lab that cannot answer from their material is decoration. A generic chatbot that sounds like a textbook is the wrong customer.

Task. Cut time-to-answer without inventing content. Answers had to come from the upload. The assistant had to be fast enough and up enough to use during real study.

Action. I shipped RAG over uploaded PDFs so Q&A was contextual. Sub-300 millisecond API responses and 99.5 percent uptime were product constraints. I added prioritized topic suggestions from previous-year-question analysis so students studied what exams actually ask. PYQ ranking is analytics, not an LLM guessing the syllabus. GitHub currently shows an AegisAI README on the GiftedBooks repo. I will not mix that product in. The resume is the source.

Result. Reading efficiency improved 35 percent. Engagement improved 50 percent. Content comprehension moved 2.5 times. Average doubt resolution went from hours to three to ten minutes. I will not invent student headcount.

Situation. Ylogx, November 2024 to October 2025. Full-stack AI BI: FastAPI, NestJS, Postgres, React, LangChain SQL RAG. The chatbot returned rows on a demo. I was not the named security team. I was the intern shipping the data path. That is when people say “filter in the UI and ship.”

Task. Own correctness and isolation for three organizational tiers, not “chatbot works on my user.” A demo that leaks is a short-term result. Uptime is part of the same ownership.

Action. I put RLS and RBAC in Postgres so generated SQL ran as the user’s database role, not as a service superuser. A JOIN could not outrun the app because the database refused the row. I put Redis on the bot path for a 35 percent drop in database latency without skipping RLS. I kept the BI app on CloudFront, ECS, Docker, CI/CD, GoDaddy into Route 53 and an ALB. I did not wait for a rewrite to make isolation real.

Result. 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, sub-210 milliseconds. I still owned the data path. I was not a manager.

-- generated SQL runs as the user role; RLS must still apply
SET ROLE tenant_user;
SELECT kpi, value FROM facts;  -- other orgs' rows do not appear

Situation. IQVIA intern, April 2026 to present, Kochi. Analysts faced more than two hundred websites and more than two-hundred-page BRDs. Stuffing the PDF invents citations. One scraper misses families of sites. Hand-written tests from a two-hundred-page requirements document do not scale.

Task. One retrieval and orchestration path that ranks sources and turns BRDs into test cases with traces, not a pile of prompts and a vector index named RAG.

Action. Deep Research: LangGraph planner, parallel Firecrawl / Bing / DuckDuckGo / Playwright, ranker on quality, recency, and agreement, then a writer. State is checkpointed so a failed scrape does not restart the run. Hybrid RAG: Azure AI Search hybrid plus semantic, GraphDB hops, LangGraph adaptive retrieval. Lexical for R-141, semantic for policy language, graph for dependencies, cite or abstain. LangSmith tracing, evals, and test-case generation made the invention checkable. I did not use Qdrant.

Result. More than two hundred websites ranked. More than two-hundred-page BRDs processed with traces and generated test cases. No extra percentage is on the resume. I will not add one.

Situation. Argus is industrial computer vision for PPE and attendance. A detector that looks fine in a notebook is not right on a floor. At 73 percent mAP you miss helmets and you fire junk boxes.

Task. Move to a number I would actually alert on, at a frame rate the stream can live with, with a log of what fired. Disconfirm the belief that 73 percent was progress.

Action. I treated 73 percent as a fail. I trained and evaluated on more than 15,000 images and iterated YOLOv9 to 89 percent mAP. I required 24 FPS. I wrote events to Postgres. I did not replace boxes with an LLM looking at frames. I did not invent Haar cascades as a first attempt. OpenCV stayed on decode and resize.

Result. Seventy-three to 89 percent mAP, 24 FPS, safety violations down 50 percent, more than twenty camera feeds, compliance doubled. GitHub argus-stream-api-server README is 404. I will not invent routes.

if (mAP < 0.89) do not scale to 20 cameras;
if (fps < 24) the model is not "right" on a live stream;
log every alert to Postgres;

Situation. Ylogx had been FastAPI, NestJS, Postgres, and SQL RAG. IQVIA needed multi-agent research, hybrid retrieval, GraphDB hops, and tracing. Tutorial-complete is not architecture.

Task. Learn enough to design the graphs: node, state, tool, and how we know a run failed, on more than two hundred sites and on two-hundred-page BRDs.

Action. I used LangGraph so a failed scrape does not restart the whole run. I used hybrid plus semantic plus GraphDB because a long BRD is not a single embedding query. I used LangSmith so curiosity had a pass/fail, including generated test cases. I discarded “RAG means one vector index.” I discarded a linear chain for Deep Research because retries and parallel tools are a graph.

Result. Two hundred plus sites ranked, two-hundred-page BRDs, evals and test-case generation. Production is Python FastAPI. Live Code is Java.

Situation. Tech Team at IEDC CUSAT. CSE at CUSAT, CGPA 8.42. 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.

Task. Raise the floor so a teammate could ship. Shared setup and the path we had already burned, not a mentor title on a slide.

Action. I shared environment setup, reviews, and “here is the failure we already hit.” 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.

Result. 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.

Situation. Argus sits on industrial safety. 73 percent mAP is a model card. It is not a standard you page a supervisor on. Alarm fatigue is the cost of a low bar. A model that is accurate in a notebook and slow on the stream is also a low bar.

Task. Make the standard three things together: the eval number, live FPS, and logged alerts. Missing any one is not almost production.

Action. More than 15,000 images. YOLOv9 from 73 to 89 percent mAP. Twenty-four FPS on the stream. Postgres logs. A path that could support more than twenty cameras. I did not scale the 73 percent model to look busy. Defects not sent down the line. The problem stays fixed because the gate is mAP plus FPS plus logs.

Result. Safety violations down 50 percent. Compliance doubled. The standard was the gate, not a paragraph in a README.

Situation. 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. The GitHub README is default Next.js. I will not invent features from it. The resume is the product.

Task. Make experiment, share, and reuse the default. Thinking small would have been one model on my laptop. Looking around the corner is CV people who are not ROS experts still prototyping in the browser.

Action. I built a 3D simulation lab for real-time experimentation. 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.

Result. Iteration time down 30 percent. More than fifty marketplace items. I will not add a go-to-market number or a user count.

Situation. Team Horizon, February to June 2024. ERC 2024’s date is fixed. Semi-autonomous Mars rover on ROS2 with student hardware. Waiting for extra sensors as the first move is delay dressed as prudence. An irreversible collision is not the kind of decision you rush.

Task. Ship perception and planning that can run. Action without fusion is just a fast video.

Action. 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. Downsampling for viz and keeping density for the local costmap was a same-week choice.

Result. 17th globally of more than eighty teams. Obstacle detection up 40 percent. Collision risk down 55 percent.

Situation. Ylogx ran BI and SQL RAG on Postgres. Every natural-language question hitting the database is both latency and money. The obvious spend is a larger instance. The intern’s temptation is to call that scale.

Task. Cut load without buying a bigger database as the first move, and without dropping isolation.

Action. 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. One routing path: GoDaddy DNS into Route 53 into an ALB. Cached answers still had to be tier-correct under RLS. A dashboard KPI that must be sub-210 milliseconds does not go through an LLM at all. I will not invent an AWS bill or a TTL.

Result. Minus 35 percent bot database latency, sub-210 milliseconds, 99.9 percent uptime, 40 percent faster reports. Thirty dashboards and plus 60 percent ops efficiency are the broader serving result, not a second spend story.

NL question
  → Redis (hot schema / repeat answer, still tier-scoped)
  → FastAPI SQL RAG → Postgres as user role (RLS)
Dashboard KPIs: API + cache, no LLM on the sub-210 ms path

Situation. The awkward sentence I had to say out loud is that a pretty lab that cannot answer from the student’s own material is decoration, and a generic chatbot that sounds like a textbook is not a teaching assistant. I will not pretend the VR surface was enough. GitHub mismatches AegisAI. Resume only.

Task. Benchmark against a real TA, not against a demo that feels smart. Ground answers. Keep the suite up. Be honest about what exams ask.

Action. I said the failure mode before we dressed it up: hallucinated citations and downtime both burn trust. I shipped contextual PDF Q&A. I held sub-300 milliseconds and 99.5 percent uptime. I used PYQ-prioritized topics. Per-user PDFs must not retrieve another student’s notes. I did not claim the suite was already a substitute for a teacher.

Result. Doubts from hours to three to ten minutes. Reading up 35 percent, engagement up 50 percent, comprehension 2.5 times. I will not claim zero incidents. I will not say the VR looked expensive therefore the work was done.

Situation. The Ylogx SQL RAG chatbot was working while bot database latency and cross-tier risk were the real product. On a happy-path account the model wrote SQL, Postgres returned rows, and a non-technical user nodded. That is the anecdote. The metrics disagreed: every natural-language turn was hitting Postgres, and a JOIN could outrun an app filter. Measuring the query path myself was not beneath the intern ticket.

Task. Separate three hypotheses: bad query shape, missing cache, and security checks in the wrong layer. Isolation for three tiers had to live where SQL actually runs. If I only sped the bot up, I might have cached a leak.

Action. I measured bot database latency instead of celebrating that SQL came back. I put RLS and RBAC in Postgres for those three tiers. 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. Same serving discipline: CloudFront, ECS, Docker, ALB.

Result. Bot database latency dropped 35 percent. Analysis productivity plus 65 percent. Reports 40 percent faster. 99.9 percent uptime. Dashboards sub-210 milliseconds. Thirty KPI dashboards, plus 60 percent ops. The deep dive was latency plus isolation, not a search-console ticket.

Situation. 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 default for RAG on PDFs is a vector index and a demo. IQVIA BRDs are more than two hundred pages. A teammate or stakeholder who wants vector-only is not being foolish. They are optimizing for a ship date. I still had to say the failure mode out loud.

Task. Argue for Hybrid plus Semantic Azure AI Search plus GraphDB plus traces. After the path is chosen, stop debating tools and instrument the path.

Action. I stated that similar embeddings are not requirement coverage. Clause IDs and dependency graphs do not fall out of cosine. I proposed hybrid plus semantic plus GraphDB plus LangGraph adaptive retrieval. Once we chose the path, I committed via LangSmith evals and test-case generation. Commit wholly means instrument the chosen path, not keep a quiet side project on FAISS.

Result. Hybrid RAG as on the resume. Two-hundred-page BRDs. Evals and test-case generation. Disagree on retrieval. Commit on evals.

Situation. ERC 2024, more than eighty international teams. I was Horizon software from February to June 2024. The date does not move because your lighting in simulation still looks clean. Dirt and dust were not in the Gazebo world we trained our eyes on. Activity would have been more bags and more slides. The result judges score is a rover that runs on the day.

Task. A field-capable perception and planning stack by that date. 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.

Action. I shipped GStreamer at 60 FPS. I took ZED 2 mapping at more than two million points per second into RViz and Gazebo. I put costmap planning and sensor fusion in front of the actuators. When the costmap lagged, the rover planned on stale occupancy. That was the ship-blocker. We downsampled for visualization and kept density for the local costmap. Public Gstreamer-UDP is a supporting webcam artifact, not a second product.

Result. We placed 17th globally of more than eighty teams. Obstacle detection up 40 percent. Collision risk down 55 percent after fusion, not after a purchase order. The run happened. The resume does not state a miss. I will not invent one to sound humble.

Situation. 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.

Task. Make the software path teachable under a fixed date. 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 subsystem.

Action. I worked as a software team member, not a hero module. I shared GStreamer, ZED, and costmap constraints so others could integrate. Same habit at IEDC and at hackathons: the person next to you can run the build. That is an environment, not a program name I will invent.

Result. 17th of more than eighty is a team result. IEDC Tech Team and eight hackathons are on the resume. I will not claim a happiness score or a mentee count.

Situation. One camera demo does not create industrial safety. Wrong alerts at more than twenty cameras scale harm as well as good. Alarm fatigue is the downstream effect. I will not invent a planet or community metric. Humility is 73 percent mAP is not coverage.

Task. Scale only a detector that passed mAP. Log what you alert. Do not multiply a 73 percent model across a site to look like coverage.

Action. 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. Video stays off the database; events go in. That is create more than you consume: an audit trail, not twenty copies of a miss. I will not invent EKS because Ylogx is the ECS story.

Result. More than twenty cameras, minus 50 percent violations, 2 times compliance, after the eval gate, not before it.

I am a software developer intern at IQVIA in Kochi, from April 2026 to the present. Two systems, not we called an LLM. LangGraph Deep Research: Firecrawl plus Bing, DuckDuckGo, and Google Playwright, rank more than two hundred websites. Hybrid RAG: Azure AI Search hybrid plus semantic, GraphDB, LangGraph adaptive retrieval on more than two-hundred-page BRDs, LangSmith evals and test-case generation. Production is Python FastAPI. Live Code is Java. Before that: Ylogx November 2024 to October 2025, 99.9 percent uptime and 40 percent faster reports. Before that: Horizon ERC 2024, 17th of more than eighty. 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.

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. You miss helmets and you fire junk boxes. Either failure trains people to ignore the alarm. I did not treat 73 percent as a starting blog metric. I trained on more than 15,000 images, iterated YOLOv9 to 89 percent mAP, required 24 FPS, then scaled to more than twenty cameras with Postgres logs. Violations down 50 percent, compliance doubled, after the gate. I do not have a confirmed manager quote. I will not invent one. Backup if they already heard Argus: bot database latency before Redis, cut 35 percent.

I do not have a named interpersonal blow-up. The conflict I will speak is retrieval. The fast path for RAG on PDFs is a vector index. Two-hundred-page BRDs have clause IDs cosine misses. I stated that similar embeddings are not coverage. I argued for hybrid plus semantic Azure AI Search plus GraphDB plus traces. Once we chose the path, I committed by instrumenting evals and test-case generation instead of reopening the tool debate every week. Disagree on retrieval. Commit on evals. If they already heard IQVIA: RLS in the database versus app-only. A JOIN leaks. The commit was three-tier RLS as a platform rule, not a UI hide.

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 wants how you flag risk. ERC was fixed. More than eighty teams show up. I sequenced 60 FPS and the costmap so the run happened. We placed 17th. 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. “We might miss ERC” in week twelve is useless. “Costmap is planning on stale occupancy; fusion is the path; extra sensor is the cut” is communication.

IQVIA Deep Research and Hybrid RAG were the POC series. One scraper is not enough — Firecrawl, Bing, DuckDuckGo, Playwright fail on different sites. Vector-only is not enough on two-hundred-page BRDs. A linear chain cannot retry or hop a requirement graph. Stuffing the PDF dies on tokens. LangSmith was the stop: 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. Playwright on every URL loses on cost; Bing snippets first, full render when the snippet is thin. Result: more than two hundred sites ranked, more than two-hundred-page BRDs with traces. No extra percentage.

Hybrid RAG on more than two-hundred-page BRDs. Stuffing invents citations. Vector-only misses clause IDs. Limited information on Deep Research: no single gold page across more than two hundred websites. I ranked on quality, recency, and agreement. I refused to answer without a retrieved span. Two high-ranked pages that contradict each other is false certainty if you take first-hit. Backup physics: Horizon field versus Gazebo. Fusion on ZED 2 plus a live costmap. I pick one and go deep. I do not list all three of IQVIA, Horizon, and Argus.

Challenge I solved: stale occupancy on Horizon — operator feed healthy, planner lying — then 60 FPS plus live costmap, 17th of more than eighty. Problem I identified: isolation living only in the Ylogx application while generated SQL could JOIN; I put three-tier RLS in the database. Simplifying: citation-gated writer plus GraphDB hop instead of stuffing two hundred pages; Redis instead of a bigger RDS. Probes on whether the first story was challenging: yes, three clocks — GStreamer latency, ROS2 callback, ZED rate. Rate-limiting decisions as story follow-ups on that IE were not the Rate Limiter LLD. Independent SDE I Rate Limiter count stays 1, not UTA.

Feedback: Argus 73 percent. Critical issue: JOIN leak and uncached NL turns on Ylogx, or stale costmap on Horizon. Beyond the task: three-tier RLS while the ticket was a chatbot demo; GStreamer at 60 FPS while a lab MJPEG would have been the designated screenshot. Didn’t know what to do next: two high-ranked pages contradict — I ranked by agreement and refused a span-less sentence instead of picking a fluent compromise. Same hour on 6806195 also had Nice Subarrays DSA. Stay ready to write Java.

ERC 2024 is a calendar date. I sequenced 60 FPS and a live costmap first. Compromise was software planning versus more hardware, not skip safety. Outside scope: GStreamer and mapping as software-team ship-blockers. Ylogx: isolation and latency were not just SQL generation. GiftedBooks backup for “not my task”: doubt hours even if the named work was VR labs. Doubts went to three to ten minutes.

If they want a team and a date: Horizon software that actually ran. Stale occupancy was the bug class. 17th of more than eighty, plus 40 percent obstacle, minus 55 percent collision. If they want a customer: GiftedBooks RAG, hours to 3–10 minutes, sub-300 milliseconds, 99.5 percent uptime. Listing three proud projects fails first. If intro already used Horizon, pick GiftedBooks.

I stay on whichever STAR they are already in. IQVIA: hybrid plus graph because vector-only misses IDs; I would freeze an eval slice earlier. Ylogx: RLS in the database because app-only fails on JOINs; denied cross-tier tests from day one; success is minus 35 percent, 99.9 percent, three tiers. Argus: do not scale 73 percent; hold out a camera; success is 89 percent mAP, 24 FPS, minus 50 percent violations. Horizon: fusion over extra hardware; time-box against collision risk earlier; success is 17th of more than eighty. Disagreement is Backbone: hybrid versus vector, or RLS-in-DB versus app-only, then commit.

I put the failure mode on the table before the library debate. Vector-only misses R-141. App-only misses a JOIN. The convince is an eval or a denied-tier fixture, not volume. Once decided, I committed wholly — traces on the chosen path, RLS as the platform rule. I will not perform a fight.