Boody — AI Nutrition Coach, Concept to Live Users in 3 Days

Boody started because every calorie tracking app I tried was the same. Scan a barcode, search a database of American foods, get a number. Nobody tracked koshari, ful medames, or hawawshi. And the coaching was always the same generic "eat less, move more" advice that ignores why people actually fail at nutrition.

I researched first principles of weight loss: physics (energy balance), biology (metabolic adaptation, protein leverage), and psychology (habit formation, decision fatigue). Documented everything in a 500+ line framework that drives every coaching decision. The AI does not wing it. Every prompt is grounded in the framework. Protein first, sleep-aware, behavior-focused, never guilt-trips.

The technical bet was self-hosting. The first model, Gemma 4 26B, hallucinated too often on food identification, so we moved to Qwen3-VL 32B, a dedicated vision-language model chosen for accuracy over speed. The pilot validates on local Apple Silicon via Ollama, the cheapest way to prove the model and the product. For production the same model self-hosts on cloud GPUs the way it should be done: vLLM for continuous batching and concurrency, RunPod Serverless so idle GPUs scale to zero, and AWS Lambda for orchestration. Users send a food photo and get back a breakdown: what the food is, estimated portions, calories, protein, and a USDA-normalized name for database lookup. No per-token API bill; at scale that is roughly $0.001 per interaction versus $0.01-0.05 for API-based competitors.

The memory system is what makes Boody feel like a real coach, not a chatbot. Four layers: SQLite for structured metrics (daily logs, meals, weight trends, consistency scores), LanceDB with nomic-embed-text for semantic search across all conversations, AI-generated weekly coach notes that capture behavioral patterns, and the last 20 messages for conversational continuity. When a user says "I had the same thing as yesterday," Boody actually knows what that was.

The data flywheel is the long-term play. Every food photo gets labeled. Every correction gets stored. User portions are learned over time. Foods that get confirmed 3+ times graduate from user-contributed to the main database. At 1,000 users, that is roughly 90,000 labeled food photos per month, building a MENA food dataset that does not exist anywhere else.

I built and launched the entire product in 3 days using Claude Code. 27 TypeScript modules, 250+ automated tests, 5 pilot users sending food photos and receiving coaching. The same architecture scales from the local pilot to the cloud GPU setup when traffic justifies it.