FDA AI Navigator

Explainable Multi-Agent Regulatory Intelligence for FDA AI Medical Devices

FDA AI Navigator is an AI-powered regulatory intelligence platform designed to help medical device builders, clinical researchers, and hospital operations leads explore and evaluate FDA-approved AI medical devices using authoritative, grounded FDA records.

Unlike a generic chatbot or simple wrapper, FDA AI Navigator combines deterministic FDA data retrieval, multi-agent reasoning, and evidence-based explanations to answer complex regulatory questions with transparent citations and verified confidence metrics.

---

🔗 Live Deployments

• 💻 Deployed Web Application: https://fda-frontend-827123939536.us-central1.run.app
• 📊 Presentation Slide Deck: https://fda-demo-827123939536.us-central1.run.app/
• 🐙 GitHub Repository: https://github.com/scrivner-solutions/FDA-AI-Devices

---

🛠️ System Architecture & Engine-Seam

FDA AI Navigator features a hybrid architecture separating deterministic data extraction from agentic reasoning:

1. SQLite Database Artifact: Seeds 1,524 AI devices, 628 companies, and key classification codes. Compiled at build-time to ensure instant, reproducible, offline-first reads.
2. Engine Seam (Storage Abstraction): The backend database access layer is decoupled from the underlying storage. While SQLite is used as the current lightweight artifact, the Engine Seam allows the backend to scale seamlessly to PostgreSQL or Cloud SQL as database sizes expand.
3. FastAPI Backend (Branching Point): Serves two distinct execution paths:
   • Deterministic APIs (No LLM): High-speed SQL query filters (/api/devices, /api/trends, /api/pathways) for 100% reproducible metrics.
   • Agentic Reasoning (Google ADK + Gemini): Executes structured cognitive tasks for complex regulatory questions.
4. Evidence Validator (Gating Layer): A deterministic Python execution layer that verifies every citation returned by the LLM against the raw SQLite database before the answer reaches the client. Unverified citations are completely discarded.

---

Architecture

The Root Agent coordinates deterministic FDA retrieval, enrichment, regulatory explanation, and evidence validation. Gemini may explain retrieved evidence but must not invent device facts.

Execution Sequence

1. Intent Agent: Parses natural language query, identifying target panel specialties, pathways, and predicates.
2. FDA Search Agent: Queries the SQLite database using deterministic SQL queries via dedicated tools (no guessing or scraping).
3. Enrichment Agent: Joins metadata, company information, and AI category tags onto the matches.
4. Regulatory Agent: Synthesizes the precedent timeline and substantial equivalence pathways.
5. Evidence Validator: Verifies all referenced K-numbers and De Novo numbers against the database and appends confidence grades.

---

📂 Repository Structure

• backend/: FastAPI backend implementing the Google ADK multi-agent framework.
• frontend/: React + Vite client interface.
• demo/: Interactive Keynote-style slide deck (accessible at demo/index.html).
• datasets/: Raw and normalized FDA seed data.
• scripts/: Production sync and data ingest scripts (run_sync.py).

---

🚀 Running Locally

Backend Setup

cd backend
pip install -r requirements.txt
PYTHONPATH=. python3 -m uvicorn main:app --port 8000

Frontend Setup

cd frontend
npm install
npm run dev

Navigate to http://localhost:5173/ in your browser.