All projects
Case study · 2024 — Present
PlanEat AI — Multi-Agent Nutrition
Fridge vision + biometric context meal negotiation
Independent builder · Independent
Personal product
Next.jsLangGraphSupabaseYOLOv8React Native
The problem
The problem
Meal planning apps demand hours of manual input. Users want an agent that sees what’s in the fridge and negotiates a plan against real constraints.
Outcomes
- Concept-to-revenue path in weeks as a consumer agent product
- Episodic RAG memory for dietary history; tool-based calorie math (LLM does not free-hand numbers)
Hard technical choices
Long dietary history saturates context — retrieve episodically instead of stuffing the window.
Deterministic nutrition calculator tool; LLM handles preference negotiation only.
Key tradeoffs
Math
Chose: Tool-enforced calorie logic
Rejected: LLM arithmetic
LLMs invent numbers; nutrition cannot be vibes.
What was hard
- Vision accuracy on messy fridges
- Trust calibration for health-adjacent advice
What I would change
- Stronger eval suite for meal constraint satisfaction
Artifacts
