Essays, papers, and system notes
Notes on delivery, process, and shipping under constraint — how the work actually gets done.
The chatbot on this site is not a demo. It is a production-grade retrieval-augmented agent — grounded answers, abstention, a review-gated evidence pipeline, versioned claims. The engineering is what an enterprise deploys on its own data. The only difference is the corpus: I built it on the one I have complete access to and authority over — my own work logs.
A framework for allocating model inference and human attention through cheap probes, measured outcomes, and recoverable allocation decisions. Each completed cycle should improve the basis on which the next commitment is made.
Could I turn my AI chat history into compact knowledge graphs without losing what I return to those conversations for? I built a harness to find out. Compression worked; reconstruction depended on what I asked for. The harness is the point; the graph is what it produced.
Replacing the header on every page of Zocdoc, across four to five codebases, under a company accessibility mandate. The work was correct long before it could land — and the fix was changing the size of the reviewable unit, then measuring each step of the rollout before widening it.