AI engineer ~ I build LLM systems that prove their work. Prefers primitives >>> frameworks

🟢 status — 2 AI systems shipped · 🚧 building a review-gated PR risk agent · writing at byraj.dev · India, remote-friendly

Upload an agreement and its amendments; get obligations extracted (owners, deadlines, penalties), an explained diff, and every change mapped to the clauses it touches. Each statement cites its source text, low-confidence items queue for human review, and every LLM call's cost is persisted per tenant.

1.0 citation validity · 0.90 extraction precision · 0.96 recall@10 FastAPI pgvector + Postgres FTS (RRF) Postgres job queue (SKIP LOCKED) AWS ECS/Fargate Terraform Next.js

Search video like it's text: ask for a moment in plain language, get the exact frame back with jump-to-timestamp links. Stage-cached ingest (FFmpeg → shot detection → faster-whisper ASR → SigLIP embeddings), fused visual + transcript retrieval over LanceDB, served as REST and MCP tools for agents.

Recall@5 0.688 vs 0.578 baseline · MRR 0.495 · p50 189 ms / p95 234 ms Python FFmpeg faster-whisper SigLIP + bge-small ONNX int8 LanceDB FastAPI MCP Fly.io

Standalone OTP timer for React Native.

• No citation, no claim. LLM output must trace to source text; invalid items get one repair pass, then drop.
• Evals before vibes. Precision/recall on frozen sets; CI fails when quality regresses.
• Budgets are product features. Per-run caps, per-tenant limits, cost persisted per call.
• Humans hold the write. Agents draft; people approve.

• Agent harnesses — tool loops, checkpointing, interrupt/resume when the model misbehaves.
• Multi-agent systems — where committees beat one loop, and where they just multiply the bill.
• Harder evals — frozen query splits, Recall@5 / MRR, regression gates instead of vibes.
• Local-first agent spend — metering and capping coding agents across vendors with a SQLite ledger and hooks.

AI/LLM — RAG (pgvector + FTS, RRF) · multimodal retrieval (SigLIP + bge-small, ONNX int8) · evals · structured outputs · LangGraph agents · MCP · LanceDB · LLM cost tracking

Backend — Python · FastAPI · Pydantic v2 · SQLAlchemy · PostgreSQL · Node.js

Frontend — React · Next.js · TypeScript · Redux Toolkit · TanStack Query · Tailwind

Infra — AWS (ECS/Fargate, RDS, S3) · Terraform · Fly.io · Docker · GitHub Actions · Sentry · CloudWatch

Full-stack at product startups (Innovaccer, Geeks Invention). Now building AI systems end to end.

Off the clock: singing and tactical shooters (Valorant, CS 🎮)