Skip to content

README.md

Latest commit

 

History

History
194 lines (150 loc) · 7.79 KB

README.md

File metadata and controls

194 lines (150 loc) · 7.79 KB

Pharma Compliance Swarm

Find the Contradictions Humans Miss

"Watch 50 adverse event reports analyzed in 2 minutes with full audit trail."

🎬 Demo Video

Pharma Compliance Demo

Script (60 seconds):

[0:00] "50 adverse event reports. FDA requires review within 15 days."
[0:10] [Agent swarm activates: Reader, Classifier, Escalator]
[0:20] [Dashboard: Processing at 25 reports/minute]
[0:30] [Alert: "Serious AE detected - Death reported - Case #AE-2024-0742"]
[0:40] [CMVK verifies: 3/3 models agree on seriousness classification]
[0:50] "50 reports. 2 minutes. 3 serious AEs found. Zero policy violations."

🚀 Quick Start (One Command)

cd examples/pharma-compliance
cp .env.example .env
docker-compose up

# Wait 30 seconds, then open:
# → http://localhost:8083  (Demo UI)
# → http://localhost:3003  (Grafana Dashboard - admin/admin)
# → http://localhost:16689 (Jaeger Traces)

📊 Live Dashboard

┌─────────────────────────────────────────┐
│ Pharma Compliance - AE Processing       │
├─────────────────────────────────────────┤
│ Reports Processed:       47             │
│ Serious AEs Found:       3              │
│ CMVK Confidence:         96.8%          │
│ Processing Time:         2.4s (avg)     │
│ Escalations:             3              │
│ Policy Violations:       0              │
└─────────────────────────────────────────┘

Overview

FDA drug applications are 100,000+ pages. One contradiction between lab reports can delay approval 6-12 months. Manual review takes weeks and misses subtle conflicts.

This demo shows how Agent OS uses Context as a Service (CAAS) and Agent VFS to perform deep document cross-referencing.

Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                     DOCUMENT CORPUS                                  │
│          50 Lab Reports + 1 IND Draft (100K+ pages)                 │
└──────────────────────────┬──────────────────────────────────────────┘
                           │ Indexed in Agent VFS
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│                     AGENT VFS                                        │
│  /agent/compliance/mem/documents/                                    │
│  ├── lab_reports/                                                    │
│  │   ├── report_001.json                                            │
│  │   └── ...                                                         │
│  └── drafts/                                                         │
│      └── ind_filing.json                                             │
└──────────────────────────┬──────────────────────────────────────────┘
                           │
            ┌──────────────┴──────────────┐
            ▼                             ▼
┌──────────────────────┐    ┌──────────────────────┐
│    WRITER AGENT      │    │  COMPLIANCE AGENT    │
│    Drafts clinical   │    │  (Adversarial)       │
│    summary           │    │  Scans for conflicts │
└──────────────────────┘    └──────────────────────┘
            │                             │
            └──────────────┬──────────────┘
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│                     CONTRADICTION REPORT                             │
│  - "Draft claims 95% efficacy, Lab Report #23 showed 89%"           │
│  - Citation: Page 42, Paragraph 3                                    │
│  - Recommendation: Update or explain variance                        │
└─────────────────────────────────────────────────────────────────────┘

Agent Types

1. Writer Agent

  • Drafts clinical summaries
  • Synthesizes data from multiple lab reports
  • Must cite sources for every claim

2. Compliance Agent (Adversarial)

  • Scans all documents for conflicts
  • Cross-references claims against source data
  • Flags contradictions with citations

Key Features

Context as a Service (CAAS)

  • 200K+ token context window (Claude 3.5)
  • Entire document corpus available for analysis
  • No information loss from chunking

Agent VFS

  • Documents stored in virtual file system
  • Standard mount points: /mem/documents/, /state/
  • Backend-agnostic (can use vector store)

Citation Linking

  • Every claim traced to source document
  • Page number + paragraph reference
  • Explainable contradictions

Policy Enforcement

  • "No hallucination" policy at kernel level
  • Self-Correcting Agent Kernel (SCAK) catches invented data
  • Must cite source for every claim

Quick Start

# Run the demo
docker-compose up

# Or run locally
pip install -e .
python demo.py

# Run with custom documents
python demo.py --reports ./my_reports/ --draft ./my_draft.pdf

# Run contradiction analysis only
python demo.py --mode contradiction_only

Demo Scenarios

Scenario 1: Efficacy Contradiction

Draft claims 95% efficacy, but Lab Report #23 shows 89%.

Scenario 2: Dosage Discrepancy

Draft recommends 10mg dose, Lab Report #7 tested up to 8mg only.

Scenario 3: Statistical Error

Draft reports p<0.001, Lab Report #15 shows p=0.03.

Scenario 4: Timeline Inconsistency

Draft claims 12-month follow-up, Lab Report #42 covers only 9 months.

Sample Output

CONTRADICTION REPORT
====================

Found 12 contradictions in 8 minutes:

1. EFFICACY MISMATCH (HIGH SEVERITY)
   - Draft: "Primary endpoint showed 95% response rate"
   - Lab Report #23, Page 42: "Response rate: 89% (95% CI: 85-93%)"
   - Recommendation: Update draft to match lab data

2. DOSAGE DISCREPANCY (MEDIUM SEVERITY)
   - Draft: "Recommended dose: 10mg daily"
   - Lab Report #7, Page 15: "Maximum tested dose: 8mg"
   - Recommendation: Add justification or adjust dose

3. STATISTICAL ERROR (HIGH SEVERITY)
   - Draft: "Statistical significance (p<0.001)"
   - Lab Report #15, Page 28: "p = 0.03"
   - Recommendation: Correct p-value in draft

... (9 more)

Metrics

Metric Human Review Agent OS
Time to Review 2 weeks 8 minutes
Contradictions Found 3 12
False Positives N/A 1
Citations Provided Partial 100%

License

MIT