What I'm focused on right now

Littman — legal intelligence. Right now I'm concentrated on the unglamorous middle of the pipeline: tightening structured extraction so the rules operate on clean facts, and calibrating confidence so the system abstains when it should instead of guessing.

FlightReady — a transparent go/no-go risk model for general aviation. The risk explorer is live; next is sharpening how the tail-risk view is explained so a pilot trusts it without needing the math.

A running bench of open questions — attribution precision, distribution-free abstention, calibration under drift, the half-life of a detector. Some have a working demo in the lab already; when one turns into something real, it graduates into a project or a note.

A short series on trust in high-stakes AI — why evidence should travel with the answer, and what it means for a system to know when to abstain.

Reranking architectures and how far you can push retrieval quality before you need a bigger model; expected-shortfall methods beyond the finance literature; and the practical edge of building with specialized agents.

Chasing every new model release, and adding features to the labs that would make them flashier but less honest. Depth over breadth.