Puddle

We built the whole AI pipeline for the Meta Model API, following Meta’s cookbook. SAM 3.1 is called over the Responses API with a point prompt built from the child’s tap (using meta_sam_parser). The returned outline steers the motorised stage. Muse Spark 1.3 is called over chat completions with a JSON schema whose enum is our checked list of pond creatures plus “unknown”, so it can’t invent a creature. It writes only one short live line about what the creature is doing; every fact comes from our sourced list. We make one SAM call per discovery and track on the device between calls, and every call has a timeout and an offline fallback.