I build LLM systems where the output has to be right before anyone acts on it - and where every number has a deterministic owner. The pattern across my work: Python owns the facts, the model investigates, explains, and cites, a verifier rejects what it cannot ground, and the honest limits are documented next to the claims.
Founder & sole engineer of NutriCart (DPIIT-recognised); MSc Business Analytics, University College Cork. Currently building the governance layer for agentic AI: Turnstile, a policy, audit and budget gateway for MCP tool calls.
| Repo | What it proves |
|---|---|
| turnstile | Agent governance. Default-deny policy engine, OIDC group→scope identity, hash-chained tamper-evident audit log, single-use human approvals, shadow-mode policy simulation - built against the 2026-07-28 MCP revision, mypy --strict, real-subprocess transport tests. |
| quarterline | A research agent that can't lie. SEC filings in, explainable research out: provenance for every number, a 10-check validation gate in front of the LLM, bounded LangGraph agent, golden-set evals. 567 tests. |
| bondlens-cmbs-surveillance | Cited financial analysis over real regulatory data. SEC EDGAR CMBS filings; deterministic tools own every figure; audited a 5/5 evaluation, found fallback text wearing real-looking citations, fixed it, re-measured. |
| investrag-studio | RAG as a measured engineering choice. 12 document formats, four chunking strategies, six vector stores, hybrid dense+BM25 with reranking, abstention when evidence is thin. Success@5 1.00, nDCG@5 0.848 on a self-built regression set. |
| altsignal-intelligence | Leakage-safe ML pipelines. Five official data APIs, immutable as-of-date snapshots, walk-forward PyTorch forecasts against baselines, n8n orchestration - found and fixed real bugs only live runs reveal. |
| india-import-atlas-data | Open data at national scale. India's US$775B import bill, sector by sector, with 209 researched venture paths - the dataset behind indiaimport.org. |
- Deterministic tools own every number. The LLM never produces a fact; it investigates with tools and cites what it used.
- Evals are part of the build, not a wrapping. Non-deterministic systems need tests too - golden sets, regression suites, adversarial cases.
- Local-first, honestly bounded. Everything runs on one machine with no paid APIs; the READMEs say what is not verified, not only what is.
- AI-assisted, agent-orchestrated development - Claude Code / coding agents as daily instruments, with specs, pinned interface contracts and per-wave verification.
Stack: Python · FastAPI · LangChain/LangGraph · MCP · PostgreSQL/DuckDB · React/TypeScript · PyTorch · Docker · pytest/ruff/mypy-strict
📩 kasarlekhraj01@gmail.com · LinkedIn · nutricart.dev · indiaimport.org