The AI Dietician

Problem

General nutrition tools assume Western food and ignore clinical constraints. A diabetic patient in Karachi eating roti, daal and rice gets nothing usable from them, and a wrong carbohydrate number is not a rounding error.

Approach

1. 01

   Four LangGraph agents — Intake, RAG Researcher, Diet Planner, Clinical Critic. The Critic is deliberately independent, because a check performed by the same context that produced the plan is worth far less.

2. 02

   Dual-store retrieval — Postgres for exact nutrition lookups, Pinecone for narrative clinical guidance. Vector similarity is the wrong tool for exact numeric facts.

3. 03

   Two-stage retrieval — a bi-encoder recall gate using bge-large-en-v1.5, then a cross-encoder reranker for precision on the shortlist.

4. 04

   The model never computes nutrition. It selects food IDs and gram amounts; a deterministic Python solver performs every arithmetic operation.

5. 05

   Hard safety refusals for pregnancy, type 1 diabetes, advanced CKD, users under 15, and underweight weight-loss requests, with handoff to a certified dietician. Under-15 users are excluded from expert contact at the API layer.