9. Resume projects

R3/R4 project deep dives. Metrics only from the Aug 2026 resume. Full chapter: study book 07. Job 10454435 still none. Labels: Resume-derived IE-asked.

GiftedBooks GitHub currently describes AegisAI — resume wins. StratifyLabs GitHub is default Next.js. InstaRecon is not a project pitch. ConvBI as a brand name is the Ylogx SQL RAG intern work — do not add Warpflow / valAgent as Aug 2026 resume bullets unless they found the repo; then label GitHub in the first sentence.

How to pitch (60 seconds)

Problem (who hurt) → architecture (boxes you shipped) → one hard part → one resume metric. Stop. Let them dive. “Walk me through your resume” is still IQVIA plus one number from Ylogx or Horizon.

IQVIA — Deep Research

Analysts faced more than two hundred websites. One scraper misses families of sites. I architected a LangGraph: planner, parallel Firecrawl / Bing / DuckDuckGo / Playwright, ranker on quality, recency, and agreement, then a writer. A failed scrape does not restart the run because state is checkpointed. Two high-ranked pages that contradict each other is the hard part — first-hit retrieval looks confident. I rank by agreement and refuse a span-less sentence. LangSmith shows empty scrapes still feeding the writer. Result: 200+ sites ranked. No extra percentage on the resume.

10x sites: more ranker, not more Playwright on every URL. Why not one Google API? Coverage gaps and rate limits.

IQVIA — Hybrid RAG / BRD test cases

Two-hundred-page BRDs. Stuffing the PDF invents citations. Vector-only misses clause IDs. I built Azure AI Search hybrid plus semantic, GraphDB hops, LangGraph adaptive retrieval, LangSmith evals, generated test cases. Lexical for R-141, semantic for policy language, graph for dependencies, cite or abstain. Hard part is test cases from the wrong section. A generated case with no overlapping span is a fail in LangSmith. I did not use Qdrant.

Why GraphDB? Requirements depend on each other; cosine does not hop. What would you do differently? Freeze an eval slice of BRD sections to expected tests before adding a fourth search tool.

Ylogx — SQL RAG + BI + AWS

Non-technical users could not pull analysis without an analyst. I shipped NestJS / Postgres / FastAPI, a LangChain SQL-RAG chatbot, three-tier RLS and RBAC in the database, Redis on the bot path, CloudFront, ECS, Docker, Route 53, ALB. SQL runs as the user role. App-only WHERE org_id fails on a JOIN. Result: +65% analysis productivity, −35% bot DB latency, 40% faster reports, 99.9% uptime, sub-210 ms, 30 KPI dashboards +60% ops.

NL → Redis (tier-scoped) → SQL RAG → Postgres as user role (RLS)
Dashboards: API + cache, no LLM on the sub-210 ms path

10x users: cache and connection pooling and RLS tests, not “we shard” — I did not shard. Why NestJS? Modular DI for role modules. Java Live Code is a separate sentence.

Horizon — ERC 2024

Semi-autonomous Mars rover, ROS2, date fixed, dirt is not Gazebo. I shipped GStreamer 60 FPS, ZED 2 at 2M+ points/s into RViz/Gazebo, costmap fusion. Hard part: stale occupancy — operator feed looks healthy while the planner lies. We downsampled for viz, kept density for the local costmap. 17th / 80+, +40% obstacle detection, −55% collision risk. Gstreamer-UDP on GitHub is a supporting webcam artifact, not a second product. This is not Argus YOLO.

Argus — PPE

Industrial cameras, PPE and attendance. 73% mAP is not a number I will page a floor on. 15k+ images, YOLOv9 to 89% mAP, 24 FPS, Postgres event logs, 20+ cameras after the gate. Violations −50%, compliance 2×. Video stays off the database. I will not invent Haar cascades or a train/test split the resume does not name. argus-stream-api-server README 404 — no invented routes.

Why not two-stage Detectron? 24 FPS across 20+ feeds. Why not an LLM on the frame? Fluency is not mAP.

GiftedBooks

VR labs plus RAG on the student’s PDF. Doubts hours → 3–10 minutes. Sub-300 ms, 99.5% uptime, +35% reading, +50% engagement, 2.5× comprehension. PYQ topic ranker is analytics, not an LLM guessing the syllabus. Resume only — not the AegisAI README.

StratifyLabs

CV SaaS: 3D lab, browser inference, marketplace 50+ models, URDF/WebGL, Gemini RAG bots in the scene. ML iteration −30%. Think Big is reusable experiment, not a TAM I will invent. GitHub README is default Next.js — resume wins.

Education / IEDC / hackathons

CUSAT CSE 8.42. IEDC Tech Team — raise the floor, no mentee count. Eight events; CodeRecet 1st; MLH Best Project; Magnathon 2.0 runner-up. Rapid ship is Bias for Action backup, not a fake architecture.