Nutrition tracking apps that rely on an LLM to estimate calories directly are prone to hallucinated numbers, which is unacceptable for a tool meant to guide real dietary decisions. The system needed a way to use a vision model's strength at recognizing food and estimating portions, without ever trusting it to produce the actual calorie or macro values. It also needed to support coaching, goal-setting grounded in established metabolic formulas, and clinical oversight, all verified by a strong automated test suite rather than live API calls.

Designed a 2-stage pipeline that strictly separates recognition from calculation. Stage 1 uses GPT purely to identify food items and estimate portion sizes in grams — it never touches a calorie or macro number. Stage 2 matches every identified food against USDA FoodData Central and computes macros directly from real scientific data (per_100g × portion_grams / 100). Shipped this across four phases: a core photo-to-macro loop with an edit/confirm screen and confidence badges feeding daily calorie/macro rings; an AI coaching agent with full meal history in context, streaming token-by-token responses, and proactive nudges based on remaining macro targets; a goals and progress system with an onboarding wizard computing TDEE via the Mifflin-St Jeor equation, auto-generated macro targets, and 7-day adherence tracking with streaks; and a clinician dashboard with role-based access, invite-based patient-clinician linking, target overrides, and clinical notes. Every phase was built spec-first — a written contract covering DB schema, Pydantic models, API shapes, UI layout, and a verification checklist — before any code was written, which is why 54 tests pass in CI with zero real API calls using mocked vision and USDA providers.

Designed a 2-stage pipeline that strictly separates recognition from calculation. Stage 1 uses GPT purely to identify food items and estimate portion sizes in grams — it never touches a calorie or macro number. Stage 2 matches every identified food against USDA FoodData Central and computes macros directly from real scientific data (per_100g × portion_grams / 100). Shipped this across four phases: a core photo-to-macro loop with an edit/confirm screen and confidence badges feeding daily calorie/macro rings; an AI coaching agent with full meal history in context, streaming token-by-token responses, and proactive nudges based on remaining macro targets; a goals and progress system with an onboarding wizard computing TDEE via the Mifflin-St Jeor equation, auto-generated macro targets, and 7-day adherence tracking with streaks; and a clinician dashboard with role-based access, invite-based patient-clinician linking, target overrides, and clinical notes. Every phase was built spec-first — a written contract covering DB schema, Pydantic models, API shapes, UI layout, and a verification checklist — before any code was written, which is why 54 tests pass in CI with zero real API calls using mocked vision and USDA providers.

Nutrition tracking apps that rely on an LLM to estimate calories directly are prone to hallucinated numbers, which is unacceptable for a tool meant to guide real dietary decisions. The system needed a way to use a vision model's strength at recognizing food and estimating portions, without ever trusting it to produce the actual calorie or macro values. It also needed to support coaching, goal-setting grounded in established metabolic formulas, and clinical oversight, all verified by a strong automated test suite rather than live API calls.