Maillard
BBQ cook plans cited to the second of source video.
What it is
Maillard is a native iOS app with a Python backend. The user picks a grill and a cut. The backend searches 2,995 transcribed cooking videos, most of them from BBQ pitmasters, and drafts a cook plan from what those cooks actually did. A fixed table of food-safety rules then corrects or blocks unsafe numbers, and in the app each step opens its source video at the cited second.
By the numbers
What it means for your team
It shows answers drawn from a business's own content library, with citations back to the exact source and a hard-coded safety layer the model cannot override, built with written specs, pass bars set in advance and independent review.
Status
Not released: it runs in the iOS Simulator and as one developer install, with no TestFlight or App Store build. The backend runs on one Mac, and the in-recipe chat still depends on a hosted model key it does not have. The full eval failed its side-dish coverage gate. Showing creator transcripts or clips publicly would need creator licences, and none are on record.
Library
An ingestion pipeline takes each video's captions, or a local Whisper transcript when captions are missing, splits the text into 57,396 chunks and embeds them locally into Postgres with pgvector.
Retrieval
The app sends the grill, the cut and the user's preferences to the backend. The search combines vector and keyword matching and uses at most three chunks from any one video.
Drafting
A model writes the plan only from the retrieved chunks. Since 2026-10-01 this step runs on the local Qwen 3.8 Splash model, with no per-call cost.
Safety gate
Plain Python rules enforce minimum pull temperatures, the 40 to 140°F danger zone with its 4-hour limit, and a cooker temperature at least 20°F above the target internal temperature. A separate citation check flags any step without a valid source clip.
Build process
Claude writes the specs and reviews the work, and builder agents implement it. A slice lands only after a clean build, passing UI tests, an adversarial review and an independent audit, and each decision is recorded in a decision log.