Multi-Agent Compliance Reviewer (Google ADK)

Track: Agents for Good — Healthcare/MedTech compliance

A multi-agent system for rapid regulatory document review in medical devices. Uses Google ADK to fan out specialist agents (requirements, risk, test, guidelines), pipeline them through a summarizer, and loop in humans for low-confidence findings. Includes built-in tools (Google Search) and custom FunctionTools.

Key Features

• Global compliance brain: Prompts are grounded in ISO 13485/14971, IEC 62304/60601/62366, EU MDR/IVDR, FDA 21 CFR Part 820/11, NIST, and DoD STIG so findings map to real-world regulators.
• Full document coverage: Handles PDF, DOCX, TXT/MD, CSV, and legacy DOC (via textract) with a hard 10MB guardrail to keep UX safe and predictable.
• Resilient orchestration: Google ADK Parallel/Sequential/Loop patterns plus persisted job state (data/jobs_state.json) to survive restarts while keeping the audit trail intact.
• Human-in-the-loop ready: Flags low confidence/agent failures for human decisions and re-summarizes with those decisions applied.
• Demo-ready UX: CLI demo, FastAPI UI for a quick visual walkthrough, structured JSON reports for scoring.

Pitch (Problem / Solution / Value)

• Problem: Regulatory reviews for medical devices are slow, error-prone, and require coordinated expertise across requirements, risk, testing, and standards.
• Solution: A Parallel + Sequential + Loop ADK workflow that ingests a document, runs specialist Gemini agents concurrently, aggregates findings, and routes low-confidence items for human approval.
• Value: Faster compliance checks, clearer action items, audit trail for regulators, and human-in-the-loop where judgment is needed.

ADK Features Demonstrated

• Multi-agent orchestration: Parallel specialist agents; Sequential summarization; Loop for human review.
• Tools: Google Search built-in tool + custom FunctionTools (nano probe, document loader).
• Sessions & state: Coordinator session state captures specialist outputs, human reviews, summaries, and audit log.
• Observability: Structured logging; audit trail in final report.

What’s Included

• src/coordinator_simple.py — Coordinator orchestrating specialist agents, summarizer, human review loop, ADK tools.
• src/agents/* — Gemini-backed specialist agents and summarizer with global regulatory context in prompts.
• demo.py — CLI demo: runs parallel analysis on a sample doc, produces JSON report, simulates human approvals.
• app.py — Optional FastAPI UI wrapper with persisted job tracking (data/jobs_state.json) and hourly cleanup.
• config/config.yaml — Defaults (model/temperature, chunk sizes). No keys stored here.
• data/sample_medical_device_requirements.txt — Demo input.
• tests/test_document_processor.py — Parser tests including 10MB guardrail and legacy DOC via textract.
• tests/test_job_store.py — Persistence and cleanup coverage for the job store.

Setup

python3 -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

Configure credentials (preferred via environment):

export GEMINI_API_KEY="your-key"          # required
export GEMINI_MODEL="gemini-2.0-flash"    # optional override (matches config default)

You can also copy .env.example to .env and fill values, then:

export $(grep -v '^#' .env | xargs)

Local runs auto-load .env via python-dotenv (app and demo); production should use real environment variables instead.

Run the CLI Demo

python demo.py

This will:

• Load data/sample_medical_device_requirements.txt
• Run 4 specialist agents in parallel
• Flag low-confidence/failed agents for human review
• Produce data/compliance_report_parallel.json
• Simulate human approvals

Run the FastAPI UI (optional)

uvicorn app:app --reload --host 0.0.0.0 --port 8000

Then open http://localhost:8000 (lightweight HTML front-end).

Screenshots

Tests

python -m pytest tests -v

Dependencies of note

• google-adk for orchestration primitives.
• textract for legacy .doc support (may require system packages depending on OS).

Notes on Models

• Default model comes from config/config.yaml (gemini-2.0-flash). Override with GEMINI_MODEL.
• If you see 404s, list available models for your key:

python - <<'PY'
import google.generativeai as genai, os
genai.configure(api_key=os.environ["GEMINI_API_KEY"])
for m in genai.list_models():
    print(m.name, m.supported_generation_methods)
PY

Pick one that supports generateContent (e.g., gemini-1.5-flash-002, gemini-2.0-flash if available).