Back to Projects
AI & ML, Multi-Agent, Full StackCompleted

Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant

An evidence-grounded culinary safety and food chemistry assistant that prevents LLM hallucinations by coupling multimodal ingredient vision with hierarchical multi-agent orchestration and canonical scientific databases (FooDB, EFSA OpenFoodTox, PubChem).

Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant

Key Details

Hierarchical Multi-Agent Architecture: Engineered a 3-tier specialist agent system with Google ADK and LiteLLM (Gemini 3.5 Flash Lite & Qwen 3.6 35B), comprising Root Orchestrator, Food Analysis Specialist, and Research Specialist.
Evidence-Grounded Hallucination Prevention: Enforced strict Tool Budgeting (maximum 4 tool calls per analysis turn) and auditable FooDB citation tags ([<Compound:FDBxxxxxx>]), guaranteeing no fabricated chemical reactions without empirical evidence.
High-Performance Knowledge Engineering & Entity Resolution: Ingested over 70,000+ food-chemical relations from FooDB and EFSA OpenFoodTox 3.0 into SQLite; developed conservative fuzzy entity matching (SequenceMatcher + token subset containment with safety margins) mapping informal ingredient names to scientific taxonomy.
Real-Time Streaming Engine: Built a low-latency Server-Sent Events (SSE) pipeline via FastAPI delivering text_delta, tool_call, and tool_result events with sub-500ms Time-to-First-Token (TTFT).
Interactive Data & Reaction Pathway Visualization: Developed a responsive Next.js 16 (React 19, TypeScript) interface dynamically rendering Mermaid.js hazard pathway diagrams, KaTeX biochemical formulas, and bilingual i18n support (EN/VI).

Highlights

Hierarchical 3-Tier Multi-Agent System via Google ADK
Evidence-Grounded Hallucination Prevention with 4-Tool Budgeting
70,000+ Chemical Relations Indexed from FooDB & EFSA OpenFoodTox
Conservative Fuzzy Entity Resolution for Scientific Taxonomy
Real-Time SSE Streaming with Dynamic Mermaid.js Reaction Pathways

Technologies

PythonFastAPIGoogle ADKNext.js 16React 19TypeScriptLiteLLMSQLiteMermaid.jsTailwind CSSServer-Sent EventsFooDB

Project Gallery

Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant screenshot 1
Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant screenshot 2
Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant screenshot 3
Hestia: Multimodal Food Chemistry & Evidence-Grounded Culinary Safety Assistant screenshot 4