BestForce

A live copilot for walk-in shops — it spots late orders and tells staff what to do. TurnPilot watches a streaming NFC queue-event feed, builds a situational model of the kitchen, and turns it into plain-language advisories with a one-tap Accept / Dismiss loop that learns. All reasoning runs on local Gemma 4 via Ollama — offline, on-device, privacy-first. No cloud calls. It is deliberately not a dashboard: the product is the advisory + accept/dismiss loop, not a wall of charts. The console reads top-to-bottom as a story — a plain-language situation line ("3 orders running late — a customer is waiting at the counter"), the live queue, then the advice. Why Gemma, on-device: detecting an overrun is just math — Gemma decides whether it's even worth interrupting you, says exactly what to do in plain language, and (with an opt-in counter camera) reads the scene with on-device vision to sense a customer actually waiting — the one signal the queue data can't capture. Those camera frames can never leave the device, so it must run locally: private by construction, and it works even when the wifi doesn't. Every Dismiss makes it less trigger-happy.