
Learning a Meaningless Signal
A few examples teach a model what a random vector in its activations means.
Open experimentFounder of a public-benefit company Local Minutes, L3C whose free index makes public municipal records searchable through an AI-powered conversational interface, now covering nearly every Vermont municipality and millions of pages of documents (localminutes.org). Responsible for every aspect of the project: data acquisition and the crawler, a RAG pipeline with embeddings in Pinecone, retrieval quality and evaluation infrastructure, speech-to-text with diarization for meeting audio, product design, UX, site reliability, monitoring, and cost instrumentation. Local Minutes helps municipal staff save time doing research, helps journalists dig into the history of issues, and helps citizens be engaged and informed in their local governments. Also built Draft Minutes (draftminutes.org), which turns a meeting recording into draft minutes with every sentence click-to-verify against the source video, keeping the clerk in control of the record.

A few examples teach a model what a random vector in its activations means.
Open experiment
Let the model control its temperature when it wants to think divergently
Open experiment
Fewer mistakes than Gemini 2.5 Flash-Lite, at 71% less cost.
Open experiment
Finding forests and new developments in Vermont with satellite embeddings.
Open experiment
Satisfying superfluous colorful swirls make managing my to-dos more fun.
Open experiment
A fisheye effect for dense data with a smooth transition at the edge, preserving magnification at the focus.
Open experiment
An attempt to approximate the style of Holger Lippmann’s Recursive Tree X.
Open experiment