About
A practitioner who still loves the craft.
I'm an AI/ML lead who builds LLM systems the unglamorous way: schema-grounded answers, approval gates before side effects, evals used as release evidence, and a monthly budget the code itself refuses to exceed.
Recent work includes ShopFloor Copilot — a manufacturing operations agent that plans investigations in plain English, writes its own guarded SQL, and pauses for a human before it touches anything — and TenQ, an equity-research copilot MVP design for answering questions from SEC filings with citations back to the exact passages.
I believe autonomy is something a system earns, feature by feature, with guardrails you can point at. The writing on this site is the build log of that belief.
Agentic Systems
Planner–tool-loop architectures with interrupts, resumable state, and hard-kill cancellation — from LangGraph graphs to custom orchestrators when the framework gets in the way.
Guarded Autonomy
LLMs propose, tools execute, humans approve. AST-whitelisted SQL, read-only execution paths, idempotent actions, and an audit row for everything.
Cloud-Native Frugality
Scale-to-zero Cloud Run, keyless auth, secrets in a vault, budget guards in the code. My systems run on hobby budgets and are engineered better for it.
Build in the Open
Every project ships with its design docs, eval harness, and a build log. If a claim isn't reproducible, it doesn't go on the page.