What is Chef Astra?
Chef Astra is a visionary project leveraging a specialized network of communicating AI agents to deliver truly personalized, adaptive, and actionable meal planning. It moves beyond static recipe databases and simple filters to function as a concierge agent - deeply integrating with your life to automate complex, recurring decisions around nutrition.
How it works
When you make a request, Chef Astra runs it through a pipeline of 12 agents. First, a safety agent screens your input. Then your request is enriched with your profile and time-of-day context. Three health experts (Dietitian, Doctor, Sports Nutritionist) run in parallel to generate guidelines. A cuisine router selects the right specialist chef, and RAG retrieves relevant recipes from a 1,077-recipe database. Finally, a master chef assembles the plan, a reviewer validates it, and a memory agent learns your preferences for next time.
The tech stack
Built with LangGraph for agent orchestration, FastAPI for the API layer, ChromaDB for vector search, and Gemini API for LLM inference. Deployed on Azure Container Apps.
Origin: Google ADK Competition
Chef Astra started as an entry for the Google Agent Development Kit (ADK) competition on Kaggle. The original version used Google's ADK framework with SequentialAgent, ParallelAgent, and AgentTool patterns. It demonstrated how multiple specialized agents could collaborate on a complex task like meal planning.
The project was then rebuilt from scratch on LangGraph for more control, adding RAG retrieval, an evaluation pipeline, fine-tuning experiments, and production deployment - growing from a competition demo into a 6-phase engineering project.
Originally built for a Kaggle competition.
Built Through 6 Phases
Phase 1: Core LangGraph rebuild (12 agents, 3 graphs)
Phase 2: RAG with ChromaDB (1,077 recipes)
Phase 3: Evaluation pipeline (97.3% avg score)
Phase 3.5: Streaming, cancellation, onboarding
Phase 4: Expanded dataset (77 to 1,077 recipes)
Phase 5: Fine-tuning experiments (QLoRA on Gemma 2)
Phase 6: Azure deployment with Gemini API
