NOTEBOOK

Free, 60-minute live session: I point my actual agent fleet at a real analytics setup or feedback pile (yours if you bring one) and narrate what it finds, including what goes wrong. RSVP by email, no signup required.

A browser-based Grand Tour validated against R's tourr package — plus a fresh FPS bench harness that already found Canvas2D's real ceiling, before the WebGL renderer meant to fix it even exists.

Seven months of running a solo AI agent fleet across three codebases — what held up, what was theater, and the worktree bug that cost real time.

A stdlib-only Python tool watching NWS + SPC data for my house and ~10 CO/WY ski areas, three alert tiers, pinged straight to my phone — no real alert has fired yet, and the post says so.

The failure modes, validation gates, and escalation rules that actually keep a one-person agent fleet from eating its own work.

Billy Joel's and Fall Out Boy's versions of the same song, plotted against real history — and the one rank-correlation number that explains why they feel so different.

Eight family LoRa radios, each with an incomplete view of the same conversation — and the dedup layer built to stitch them into one timeline.

A proof-of-concept PWA that hides your town under game-style fog until you actually walk there. Core mechanic works; everything past that is paused.