Aaron Trachtman

AI Dogs is a mobile-first web app that helps shelter volunteers prioritize dog walks via AI whiteboard extraction. Volunteers photograph each kennel whiteboard—a handwritten grid of every dog's daily outings—and get a ranked walk list in seconds, filling in as dogs are read.

At its core is a split between perception and reasoning. The whole board goes to Qwen 3.7 Plus on Fireworks in a call, but the model's only job is to transcribe what it literally sees. Every interpretation lives in deterministic, unit-tested TypeScript instead: handwritten times to 24-hour format, activity-code normalization, time ranges, and distinguishing a "walk-next" arrow from a "still out" note. Treating extraction as perception rather than reasoning lets the model run with reasoning disabled (around 4x faster), while the rules that matter stay versioned and testable, not buried in a prompt. Results stream back incrementally as NDJSON.

It's built on Next.js 16 with PIN-gated middleware, rate limiting, and fire-and-forget Supabase logging. I wrote an eval harness that scores the model against hand-labeled boards by precision and recall to iterate on the prompt. It's deployed on DigitalOcean via Docker Compose with Caddy SSL and GitHub Actions CI/CD.