Notes from the seat.
Short engineering notes from Cleared FDE™ practice, the design decisions, refusals, and small fixes behind AI that has to work inside real missions. Everything here is public-release and synthetic-data only; the interesting parts are the decisions, not the data.
Latest first.
SBIR Phase III, explained for contracting officers and program offices: what 15 U.S.C. 638(r) allows, what the file needs, the data rights that come with it, and why the authority is so rarely used.
6-minute read Field Note 003 · September 2026 Before ARES arrivesThe Air Force has chosen its next system of record for aircrew scheduling, training management, and stan/eval. What a flying wing controls in the twelve months before migration, and where a decision layer fits above a system of record.
5-minute read Field Note 002 · August 2026 Serving live AI on a public site without burning tokensBots and bored visitors will happily spend your inference budget. The design that keeps a public AI demo honest: a small fast model, hard caps, and a deterministic fallback that never embarrasses you.
4-minute read Field Note 001 · August 2026 Your aircraft-readiness AI doesn't need your SIEMMost mission-fit tools flatter every input. Ours started doing it too, recommending security-event pipelines for a flying-schedule problem, and the fix taught us what "credible AI" actually means in front of a mission owner.
4-minute readBring the workflow these notes remind you of.
Thirty minutes with a technical lead. You leave with a candidate architecture and a clear go/no-go, not a follow-up deck.