ai/ml engineer · backend systems · agents & retrieval
i build ai systems that can be tested, checked, and debugged.
i work on the parts that make these systems reliable: execution, retrieval, evaluation, and handling failures.
insight — deterministic analytics agent
turns natural-language questions into analysis plans and runs calculations through pandas, scipy, and a read-only duckdb sandbox. independently recomputes key aggregates instead of relying on the model's arithmetic. includes an mcp server and evaluation suite.
k-state — corrective retrieval-augmented generation
combines faiss, bm25, and reciprocal rank fusion. checks retrieved evidence before generation, rewriting queries or falling back to web search when local context is weak. includes sentence-level filtering, a fastapi layer, redis caching, and golden-set evals.
heimdall — automated github code review agent
a fastapi webhook service that verifies github signatures, fetches commit diffs, and posts structured review feedback. processes commits independently so a failure in one review doesn't block the rest of a push.
engine — multi-agent drone defense simulation
built on gym-pybullet-drones. combines multi-agent target assignment, kalman-filter tracking, proportional-navigation guidance, and behavior-tree decisions. uses seeded experiments and monte carlo evaluation to compare coordination strategies and find problems.
- light-rag — grounded generation with claim-level evidence grading, citation verification, and reproducible evaluation.
- languages: python, sql, bash
- backend & data: fastapi, postgresql, duckdb, sqlalchemy, pydantic, docker
- retrieval & orchestration: faiss, chromadb, bm25, langgraph
recognition · 1st place nationwide at smart india hackathon 2025 for a drone-defense simulation developed for bharat electronics limited (bel).