04
I teach this too
// @LLMImplementationHands-on LLM engineering — RAG, agents, fine-tuning, evaluation
SubscribeMost agents just sound right. I build the ones that are right — guardrails in the tool layer, human approval where it matters, and evals that grade the database, not just the wording.
Six years building AI that ships — banking NL2SQL, then GenAI for regulated finance, now agents I run end to end. The model is the easy part. I sweat the rest: grounding, tool boundaries, evals, and owning the system after launch.
One project shows the reliability pattern companies need. One proves I can ship a real product end to end.
Six years, two phases — ~4.5 in industry, ~2 building independently. Today: live AI products, most recently Jupiter, alongside hands-on technical education and collaboration on fine-tuning and agent workflows.
Before that, ~4.5 years as NLP lead at Aunalytics, a data & analytics company — GenAI for regulated community banking: Text-to-SQL, domain fine-tuning, tool routing, retrieval, shipped where compliance wasn't optional. M.A. Statistics, Columbia · ECML PKDD 2021 (Springer LNCS).
I stress-test your agent for the ways they actually fail — tool misuse, policy leaks, fabricated confirmations — and hand back a prioritized fix plan.
A focused agent around one real business workflow — support, retail, internal ops, analytics — with tool boundaries and human approval where it matters.
Test cases, verifier facts, tool-path checks, LLM-judge rubrics, and state-based outcomes — so agent failures become measurable instead of vibes.
Hands-on LLM engineering — RAG, agents, fine-tuning, evaluation
SubscribeWhether you're shipping an agent, auditing one, or just want to compare notes on doing it right — I'm happy to talk.