AI-powered EU AI Act compliance agent. Classify AI systems by risk level, identify compliance gaps, and generate required documentation -- in minutes, not months.
#eu-ai-act #compliance #rag #langchain #langgraph #ai-agents #fastapi #nextjs #legal-tech #regtech
Every company deploying AI in the EU must comply with the EU AI Act by August 2, 2026. Non-compliance carries fines up to EUR 35 million or 7% of global revenue.
- 144 pages of dense legal text (113 articles, 13 annexes)
- Over 50% of organizations lack systematic inventories of their AI systems
- Current solution: hire consultants at EUR 50K-500K per engagement
- No automated tooling exists for EU AI Act compliance
ComplyOS is an AI compliance agent that automates what takes consultants months:
Register AI system --> Classify risk --> Analyze gaps --> Generate documentation
(30 sec) (15 sec) (20 sec) (30 sec)
graph TB
subgraph Frontend["Frontend (Vercel)"]
A[Next.js 16 Dashboard] --> B[System Registration]
A --> C[Classification Results]
A --> D[Gap Analysis]
A --> E[Compliance Chat]
end
subgraph Backend["Backend (Render)"]
F[FastAPI] --> G[LangGraph Orchestrator]
G --> H["Risk Classifier Agent"]
G --> I["Gap Analyzer Agent"]
G --> J["Doc Generator Agent"]
G --> K["Chat Agent"]
end
subgraph Data["Data Layer"]
L[(SQLite)] --> M[AI Systems]
L --> N[Assessments]
L --> O[Audit Trail]
P[(ChromaDB)] --> Q[595 Legal Chunks]
end
Frontend -->|REST API| Backend
H --> P
I --> P
J --> P
K --> P
F --> L
graph LR
INPUT["AI System\nDescription"] --> CLASSIFY["Risk Classifier\n(RAG + LLM)"]
CLASSIFY -->|High Risk| ANALYZE["Gap Analyzer\n(Articles 9-15)"]
CLASSIFY -->|Limited/Minimal| REPORT["Quick Report"]
ANALYZE --> GENERATE["Doc Generator\n(Article 11)"]
GENERATE --> OUTPUT["Compliance\nPackage"]
REPORT --> OUTPUT
style CLASSIFY fill:#1e40af,color:#fff
style ANALYZE fill:#dc2626,color:#fff
style GENERATE fill:#16a34a,color:#fff
Tested against expert-validated AI system descriptions across all EU AI Act risk categories:
xychart-beta
title "Classification Accuracy by Risk Level"
x-axis ["Unacceptable", "High Risk", "Limited", "Minimal", "Overall"]
y-axis "Accuracy %" 0 --> 100
bar [100, 100, 100, 100, 100]
| Test Case | Expected | Predicted | Confidence | Result |
|---|---|---|---|---|
| Resume Screener (HR) | High | High | 95% | PASS |
| Customer Chatbot | Limited | Limited | 90% | PASS |
| Social Scoring System | Unacceptable | Unacceptable | 100% | PASS |
| Email Spam Filter | Minimal | Minimal | 90% | PASS |
| Student Exam Grader | High | High | 95% | PASS |
10/10 (100%) accuracy on benchmark. 50 test cases written across all risk levels and Annex III categories.
xychart-beta
title "E2E Test Suite — 15/15 Passing"
x-axis ["Health", "Dashboard", "CRUD", "Classify", "Gap Analysis", "Ontology", "Chat", "Doc Gen", "Verify"]
y-axis "Tests" 0 --> 3
bar [1, 1, 3, 2, 1, 3, 1, 1, 2]
| Category | Tests | Status | Avg Time |
|---|---|---|---|
| Infrastructure | 1 | PASS | 0.7s |
| Dashboard | 1 | PASS | 0.2s |
| System CRUD | 3 | PASS | 0.2s |
| Classification (LLM) | 2 | PASS | 5.7s |
| Gap Analysis (7 LLM calls) | 1 | PASS | 82s |
| Ontology API | 3 | PASS | 0.2s |
| Compliance Chat (LLM) | 1 | PASS | 11.6s |
| Document Generation (LLM) | 1 | PASS | 58s |
| Verification | 2 | PASS | 0.2s |
| Total | 15 | 100% | — |
pie title EU AI Act Corpus (595 chunks)
"Articles" : 592
"Annexes" : 699
"Key Recitals" : 224
| Source | Chunks | Coverage |
|---|---|---|
| EU AI Act Articles 1-113 | 592 | Full text from EUR-Lex |
| Annexes I-XIII | 699 | Complete annex text |
| Key Recitals | 224 | Recitals with interpretive guidance |
| Total | 1,515 raw / 595 deduplicated | Full regulation |
graph LR
subgraph Frontend
NJ[Next.js 16] --> TS[TypeScript]
NJ --> TW[Tailwind CSS]
NJ --> LR[Lucide Icons]
end
subgraph Backend
FA[FastAPI] --> LC[LangChain]
FA --> LG[LangGraph]
FA --> PY[Pydantic]
end
subgraph AI
CL[Claude API] --> RAG[RAG Pipeline]
RAG --> CH[ChromaDB]
end
subgraph Infra
VC[Vercel] --> FE[Frontend CDN]
RN[Render] --> BE[Backend Docker]
SQ[SQLite] --> DB[Persistence]
end
| Layer | Technology | Purpose |
|---|---|---|
| Frontend | Next.js 16, TypeScript, Tailwind CSS | Dashboard, system registration, results UI |
| Backend | FastAPI, Python 3.11 | REST API, agent orchestration |
| AI/RAG | LangChain, LangGraph, ChromaDB | Multi-agent pipeline, vector search over EU AI Act |
| LLM | Claude API (Anthropic) | Legal reasoning, classification, document generation |
| Database | SQLite | System registry, assessments, audit trail |
| Hosting | Vercel (frontend), Render (backend) | Production deployment |
- 6 metric cards: total systems, high-risk count, compliance score, deadline countdown, critical gaps, compliant systems
- Getting started guide with 3-step onboarding flow
- Register AI systems with natural language descriptions
- Automatic EU AI Act risk classification (Unacceptable / High / Limited / Minimal)
- Confidence scoring with cited legal articles and reasoning
- Annex III category detection (8 categories: biometrics, infrastructure, education, employment, services, law enforcement, migration, justice)
- Article 9-15 compliance assessment for high-risk systems
- Severity-ranked gaps (Critical / Major / Minor)
- Priority action list with remediation steps
- Estimated effort per gap
- Auto-generated Article 11 technical documentation
- Compliance package with risk assessment and remediation plan
- Copy to clipboard and download as Markdown
- RAG-powered Q&A over the full EU AI Act
- Suggested questions for quick start
- Cited article references in every response
- Node.js 18+
- Python 3.10+
- Anthropic API key (get one here)
# Clone
git clone https://github.com/soneeee22000/ComplyOS.git
cd ComplyOS
# Backend
cd backend
python -m venv .venv
source .venv/bin/activate # or .venv\Scripts\activate on Windows
pip install -r requirements.txt
cp .env.example .env # Add your ANTHROPIC_API_KEY
# Fetch and ingest the full EU AI Act
python -m app.services.fetch_eu_ai_act
python -m app.services.ingest_ai_act
# Start backend
uvicorn app.main:app --reload --port 8000
# Frontend (new terminal)
cd ../frontend
npm install
echo "NEXT_PUBLIC_API_URL=http://localhost:8000/api" > .env.local
npm run dev| Method | Endpoint | Description |
|---|---|---|
POST |
/api/systems |
Register a new AI system |
GET |
/api/systems |
List all registered systems |
POST |
/api/systems/{id}/classify |
Run EU AI Act risk classification |
POST |
/api/systems/{id}/analyze |
Run compliance gap analysis |
POST |
/api/systems/{id}/generate-docs |
Generate compliance documentation |
POST |
/api/chat |
Ask EU AI Act compliance questions |
GET |
/api/dashboard |
Get aggregate compliance metrics |
GET |
/health |
Health check |
Full Swagger docs: complyos.onrender.com/docs
ComplyOS/
├── frontend/ # Next.js 16 + TypeScript + Tailwind
│ └── src/
│ ├── app/ # Pages (dashboard, systems, chat)
│ ├── components/ # UI components (sidebar, cards, panels)
│ └── lib/ # API client
├── backend/ # FastAPI + LangChain + LangGraph
│ └── app/
│ ├── agents/ # LangGraph agents (classifier, gap analyzer, doc generator, chat)
│ ├── api/ # REST API routes
│ ├── core/ # Configuration
│ ├── models/ # Pydantic schemas
│ ├── services/ # RAG, database, ingestion
│ └── data/ # EU AI Act chunks (JSON), ChromaDB, SQLite
├── PRD.md # Product Requirements Document (v2)
├── CLAUDE.md # AI assistant context
├── Dockerfile # Backend container
└── LICENSE # MIT
- Full EU AI Act text ingestion (595 chunks from EUR-Lex)
- Multi-agent classification pipeline (LangGraph)
- Gap analysis with severity ranking
- Compliance document generation with rich markdown + PDF export
- RAG-powered compliance chat with formatted responses
- SQLite persistence with audit trail
- Production deployment (Vercel + Render)
- Compliance ontology (7 articles, 25 sub-requirements, 80+ verification criteria)
- Ontology-guided gap analysis (structured per sub-requirement)
- Requirement tree UI with multi-level accordion
- E2E test suite (15/15 passing on production)
- Classification benchmark (50 test cases, 100% accuracy on initial run)
- Landing page with live classification examples
- Demo data seeding (survives ephemeral deploys)
- Evidence-based assessment (document upload + parsing)
- Validated benchmark (100+ expert-reviewed test cases)
- CNIL / France-specific guidance integration
- Multi-jurisdiction support (DE, ES, NL, IT)
- DOCX / PDF export
- CI/CD integration API
Built for the VivaTech 2026 Startup Challenges (ManpowerGroup and KPMG tracks). ComplyOS addresses the EUR 17B EU AI Act compliance market -- the largest AI regulation in history with a hard deadline of August 2, 2026.
Comparable: OneTrust reached $5.3B valuation from GDPR compliance tooling. The EU AI Act is broader, fines are larger, and AI adoption is accelerating.
Pyae Sone Kyaw (Seon) -- AI Engineer based in Paris