Aditya Painuli

About

I build production AI systems that don't hallucinate: the grounding, retrieval, and evaluation layers that make an LLMs output trustworthy and traceable to real data. Most team ships an AI feature that demos well, then watch it get the hard question wrong in the field. Multi-step reasoning, reconciling figures across tables, confusing one source for another. I work on the layer in the middle that fixes that, and I measure it, because you can't call a system grounded without proving it. Background: full-stack and AI engineer. I've taken products from zero to user-facing at 4+ startups, owning the build from concept to launch, with depth in retrieval, LLM evaluation, Postgres/pgvector, and agent infrastructure. Recent: I built an open benchmark on SEC filings measuring exactly where a typical RAG pipeline breaks and how much each grounding technique recovers. It took the hard tier from 83% to 96% and cut hallucination from 11% to 2%. Repo and writeup featured.

Experience

Education