🥗 MealGraph — Nutrition Multi-Agent System

A LangGraph + Gemini orchestrator: a Coach delegates to a Medical Assessment specialist (deterministic clinical formulas + LLM interpretation) and a Planner that combines grounded web search with a PuLP linear-program meal solver and runs its own post-solve safety check (allergy / calorie / macro tolerances). Bring your own Gemini API keys.

1. Setup
Coach model
Workers (Medical / Planner) model
Tools (WebSearch) model

2. Your profile
Sex
Activity
Goal

Active agents

  • CoachAgent (orchestrator) — Central orchestrator: turns user intent into a workflow of agent/tool calls.
  • MedicalAssessmentAgent (specialist:medical) — Produces evidence-based medical assessment + clinical flags + calculations.
  • PlannerAgent (specialist:nutrition_planning) — Personalised meal plans constrained by the medical assessment. Runs an internal deterministic check (allergy / calorie / macro tolerances) after the LP solver and self-revises up to twice before returning.

About: This demo runs a 3-agent system (Coach, Medical Assessment, Planner) on top of two safe-by-construction tools (PuLP QuantitiesFinder, Gemini-grounded WebSearchTool). The Planner runs its own deterministic check (allergy / calorie / macro tolerances) after the LP solver and self-revises up to twice. The Coach then does an LLM-graded self-review (medical flag respect, citation presence, cultural fit) before composing. See the GitHub repo for the full architecture writeup.