The production app sends a meal photo to Groq llama-4-scout vision and returns a carb/calorie range with a confidence level, runs a 3-Question Protocol to tighten it, and shows a verify-before-dosing safety banner. This demo fires a real Groq text call per dish and runs an honest eval vs. USDA-derived ground truth.
Dish name + ingredients — what vision sees in production. Here we send a text description (no image in this demo).
groq llama-3.3-70b estimates carb/calorie range from dish name; result is compared to labeled Nutrition5k ground truth, computing real MAPE.
Live model estimate (tagged honestly) + carb MAPE + hit-rate vs. ground truth. Production uses llama-4-scout vision on the actual photo.
Ground-truth carbs/calories are USDA-derived Nutrition5k values. The live estimate is a real Groq text call using llama-3.3-70b-versatile; production uses llama-4-scout on the actual photo. Rate limit: 8 req/min — batch sequences calls automatically. When offline, falls back to the bundled reference estimate (labeled).