FDA Watch & Digest

Automated FDA regulatory monitoring system that uses the Federal Register API and eCFR API to detect regulatory changes and generate AI-powered summaries for quick impact assessment. This project was part of my Internship at Accelerated Biosciences in Taiwan.

Overview

Problem: Manually monitoring FDA regulatory publications and CFR regulation changes is time-consuming and risks missing important updates that could impact business operations.

Solution: Automated tool that monitors both Federal Register announcements and CFR regulation text via official government APIs, detects changes, cross-references them, and provides AI-generated summaries with email notifications.

Key Features

Federal Register Monitoring

• Automated Detection: Weekly checks of Federal Register for FDA publications
• Intelligent Filtering: 3-stage hybrid system (keywords → scoring → AI review)
• Topic Focused: Filters for cell therapy, gene therapy, stem cells, exosomes, secretomes
• 4-Tier Relevance: HIGH/MEDIUM/LOW/MINIMAL priority classification
• AI Summarization: Claude-powered summaries with key points and action items

CFR Regulation Monitoring

• Snapshot System: Periodic snapshots of 21 CFR Parts 1271, 600, 610, 312, 601
• Change Detection: Automatically detects ADDED, MODIFIED, REMOVED, RENAMED sections
• Diff Analysis: Shows before/after text comparison
• AI Impact Analysis: Relevance scoring (0-100) with detailed impact assessment
• Cross-Referencing: Links CFR changes to Federal Register documents
• Timeline Building: Shows regulatory progression (PRORULE → RULE → CFR change)
• Classification: Identifies expected vs unexpected changes

Integration & Delivery

• Interactive Workflow: Single fda workflow command for complete monitoring
• Email Digests: Combined FR documents + CFR changes with HTML rendering
• No Web Scraping: Uses official government APIs (Federal Register + eCFR)
• Historical Tracking: SQLite database maintains complete regulatory history
• Unified CLI: Single command-line tool for all operations

Usage

Primary Command: Interactive Workflow

The recommended way to use FDA Watch & Digest is through the interactive workflow:

python fda.py workflow

What it does:

• Step 1: Federal Register Detection

  • Asks: “How many days back should I check?” (default: 7 days)
  • Fetches documents from Federal Register API
  • Applies 3-stage relevance filtering (keywords → scoring → AI review)
  • Shows top HIGH priority documents
  • Auto-saves to database

• Step 2: CFR Regulation Monitoring

  • Creates snapshots of monitored CFR parts (1271, 600, 610, 312, 601)
  • Detects changes between snapshots
  • Runs AI analysis on detected changes
  • Reports relevance distribution

• Step 3: AI Summarization

  • Shows document counts (HIGH + MEDIUM relevance)
  • Displays estimated API cost
  • Asks: “Generate AI summaries? (Y/n)”
  • Processes documents with progress indicator
  • Generates 2-paragraph summaries (context + impact, no action items)

• Step 4: Email Delivery

  • Generates email digest with FR documents + CFR changes
  • Shows counts by relevance level
  • Asks: “Preview / Send / Skip?”
  • Opens preview in browser if requested
  • Sends email if confirmed (even if no changes)

• Summary: Shows what was accomplished (documents saved, CFR changes detected, summaries generated, email sent)

Quick Mode (for automation/cron):

python fda.py workflow --quick

• Uses default 7 days lookback
• Auto-summarizes HIGH + MEDIUM documents
• Auto-sends email (even if no updates)
• No interactive prompts

Individual Commands

For granular control, use individual commands:

# Document detection
python fda.py run --days 60          # Detect and save new documents
 
# Testing & validation
python fda.py filter --days 60       # Test filtering system (with AI)
python fda.py filter --no-ai         # Test without AI review (faster)
python fda.py test                   # Comprehensive system test
python fda.py smtp                   # Test SMTP connection only
 
# AI Summarization
python fda.py summarize              # Summarize HIGH & MEDIUM docs (recommended)
python fda.py summarize --all        # Summarize all relevance levels
python fda.py summarize --limit 5    # Limit to 5 documents
python fda.py summarize --relevance HIGH  # Only HIGH relevance
 
# Email operations
python fda.py preview                # Preview email in browser
python fda.py send                   # Send email digest
 
# Database viewing
python fda.py show                   # View all documents
python fda.py show --relevance HIGH  # Filter by relevance
python fda.py show --type RULE       # Filter by document type
python fda.py show --limit 10        # Limit results
 
# CFR Monitoring (NEW)
python fda.py cfr snapshot           # Create CFR snapshots for all monitored parts
python fda.py cfr snapshot --parts 1271  # Snapshot specific part only
python fda.py cfr detect             # Detect changes between snapshots
python fda.py cfr analyze            # Run AI analysis on detected changes
python fda.py cfr analyze --limit 10 # Analyze up to 10 changes
python fda.py cfr show               # View recent CFR changes
python fda.py cfr show --part 1271   # Show changes for specific part
python fda.py cfr show --relevance HIGH  # Filter by relevance
python fda.py cfr show --verbose     # Show detailed analysis
 
# Help
python fda.py --help                 # Show all commands
python fda.py workflow --help        # Help for specific command

Filtering System

The system uses a 3-stage hybrid approach to identify relevant documents:

Stage 1: Keyword Pre-screening

• Searches entire document (title + abstract + body text)
• Case-insensitive matching
• Keywords: cell therapy, gene therapy, stem cells, exosomes, secretomes, regenerative medicine
• Documents without keywords are rejected and not stored

Stage 2: Weighted Relevance Scoring

• Title matches: 10 points per keyword
• Abstract matches: 5 points per keyword
• Body matches: 1 point per keyword
• Document type bonuses:
  • RULE: +3 points (critical regulations)
  • PRORULE: +2 points (proposed changes)
  • NOTICE: +0 points

Stage 3: AI Validation (Claude)

• Only reviews borderline cases (MEDIUM/LOW relevance)
• HIGH relevance documents (≥15 pts) skip AI review (already clearly relevant)
• MINIMAL relevance documents (1-4 pts) skip AI review (clearly not relevant)
• Validates actual relevance vs. keyword spam

4-Tier Classification

• HIGH (≥15 pts): Direct regulatory impact
• MEDIUM (10-14 pts): May affect compliance
• LOW (5-9 pts): Potentially relevant
• MINIMAL (1-4 pts): Brief mentions
• REJECTED (0 pts): No keywords - not stored

AI Summarization

Uses Claude Haiku 4.5 to generate concise, regulatory-focused summaries.

Summary Structure

Each AI-generated summary includes:

• KEY POINTS: 2-3 bullet points highlighting the most important takeaways
• SUMMARY: Exactly 2 paragraphs:
  • First paragraph: What the regulation/guidance is about and why it was issued (context and background)
  • Second paragraph: Regulatory impact on cell/gene therapy companies, what changes, deadlines, and comment periods Note: Summaries focus on explaining the regulation itself, not prescribing action items. They provide regulatory context and impact assessment, letting recipients determine appropriate responses.

Cost Optimization

• Default: Only HIGH and MEDIUM relevance documents are summarized
• Rationale: Highest regulatory impact documents require detailed analysis
• Estimated cost: ~$0.01-0.03 per summary (Claude Haiku 4.5)
• Typical usage: 2-3 summaries per week = ~$0.10-0.30/month

CFR Regulation Monitoring

The system now monitors actual CFR regulation text changes, not just Federal Register announcements.

How It Works

1. Snapshot Creation (fda cfr snapshot)

   • Fetches current text of monitored CFR parts from eCFR API
   • Creates SHA-256 hash of content
   • Stores in database with date
   • Monitored parts: 21 CFR 1271 (HCT/Ps), 600/610 (Biologics), 312 (INDs), 601 (Licensing)

2. Change Detection (fda cfr detect)

   • Compares two most recent snapshots for each part
   • Identifies ADDED, MODIFIED, REMOVED, RENAMED sections
   • Generates before/after text comparison
   • Calculates similarity scores

3. AI Analysis (fda cfr analyze)

   • Relevance scoring (0-100) based on cell/gene therapy impact
   • Detailed impact analysis (2,000+ characters)
   • Compliance action recommendations (4,000+ characters)
   • Classification by relevance level (HIGH/MEDIUM/LOW/MINIMAL)

4. Cross-Referencing (automatic)

   • Links CFR changes to Federal Register documents
   • Multi-strategy matching (direct references, keywords, part-level)
   • Builds timelines showing regulatory progression
   • Classifies as expected (announced in FR) or unexpected (silent change)

CFR Change Data

Each detected CFR change includes:

• Citation: e.g., “21 CFR § 1271.90”
• Change Type: ADDED, MODIFIED, REMOVED, or RENAMED
• Before/After Text: Full text comparison
• Diff Summary: Brief description of what changed
• Relevance Score: 0-100 (AI-generated)
• Impact Analysis: Regulatory implications and affected companies
• Compliance Notes: Specific actions required, deadlines
• Related Documents: Federal Register documents that announced the change
• Timeline: Chronological progression (PRORULE → RULE → CFR change)
• Classification: Expected (announced) vs Unexpected (silent)

Email Integration

CFR changes are automatically included in email digests:

• Displayed in separate section after Federal Register documents
• Color-coded by change type (green=ADDED, yellow=MODIFIED, red=REMOVED)
• Collapsible before/after text sections
• Expandable AI analysis and compliance notes
• Related Federal Register documents listed with confidence levels

Technical Architecture

Data Flow

Federal Register API
        │
        ▼
┌──────────────────────────┐
│   Change Detector        │  Compares with database
│   (detect new docs)      │  Identifies what's new
└──────────────────────────┘
        │
        ▼
┌──────────────────────────┐
│  Relevance Filter        │  3-stage hybrid system:
│  (3-stage hybrid)        │  1. Keywords  
│                          │  2. Scoring
│                          │  3. AI validation
└──────────────────────────┘
        │
        ▼
┌──────────────────────────┐
│   SQLite Database        │  Stores relevant docs
│   (persistent storage)   │  Tracks relevance scores
└──────────────────────────┘
        │
        ▼
┌──────────────────────────┐
│  Document Summarizer     │  Claude AI generates
│  (Claude AI)             │  actionable summaries
└──────────────────────────┘
        │
        ▼
┌──────────────────────────┐
│  Email Notifier          │  Sends priority-based
│  (SMTP)                  │  digest to stakeholders
└──────────────────────────┘

Technology Stack

• Language: Python 3.11+
• APIs: Federal Register API (no auth required)
• AI: Anthropic Claude Haiku 4.5
• Database: SQLite
• Email: SMTP (Gmail or custom)
• HTTP Client: requests
• Scheduling: Cron or Python schedule (optional)

Development Status

Current Phase: Production Ready ✅

Completed Features

• Federal Register API integration
• 3-stage relevance filtering system
• Change detection and deduplication
• SQLite database with relevance tracking
• AI-powered summarization (Claude Haiku 4.5)
• Email notifications (HTML + plain text)
• Interactive workflow orchestration
• Comprehensive CLI tool
• System testing and validation

Roadmap

• Automated scheduling (cron integration)
• eCFR API integration (monitor regulation text changes)
• Web dashboard for viewing history
• Multi-user support with preferences
• Advanced analytics and trend analysis

Why This Approach?

Advantages over web scraping:

• ✅ Reliable - Official government API, well-maintained
• ✅ Free - No authentication or API keys required
• ✅ Structured - Clean JSON responses, easy to parse
• ✅ Complete - Access to historical data and metadata
• ✅ Fast - No HTML parsing or page rendering
• ✅ Legal - Using official public APIs as intended

Regulatory monitoring best practices:

• Federal Register is the authoritative source for new regulations
• Proposed rules give advance warning before final implementation
• This is where regulatory professionals look first

---

Last Updated: 2025-11-17 Status: Production Ready Version: 1.0

---

index