AI Engineer. I build LLM systems that are evaluated, not demoed — and I publish the numbers when they get worse.
sahilch.vercel.app · LinkedIn · sahilch7359@gmail.com
DataChat — agentic NL-to-SQL analytics. LangGraph plan → retrieve → guardrail → execute → verify → explain. RAG-to-schema grounding, AST SQL guardrail, read-only least-privilege role, MLflow tracing, and a 26-case golden set gating every PR. Live · LangGraph, FastAPI, Postgres + pgvector, Redis, Next.js
Quorum — supervisor agent for pull-request review over the GitHub MCP server. Hybrid retrieval (dense + BM25 + RRF); a finding is either cited to retrieved code or dropped. Nothing posts without human approval. Live · LangGraph, MCP client + server, FastAPI, pgvector
ITR6-t2sql — plain-English querying of Indian corporate tax filings. The engineering is the safety layer, not the SQL. Live · 129 statutory line items · 2 independent safety gates · 55 tests
Electricity Forecasting — 33 years of monthly US utility production. SARIMA at MASE 0.959, with the naive baseline left in the results table so the model has to earn its complexity. Live · statsmodels, SARIMA, Holt-Winters
Yardstick — an agent that trains candidate models on tabular data and reports honestly how good they are: automated leakage detection, every model scored against a trivial baseline, and metric choice justified rather than defaulted.
Platform & Data Engineer at TCS — enterprise GenAI framework, and billing-domain ETL on AWS. Retrieval quality is a data quality problem; most of what breaks in a RAG system breaks in the ingestion layer.