The Infrastructure Standard for AI Compliance on Solana

🌟 Pitch Deck | 🎥 Live Demo Video

AnnexIV-Solana-Core is an infrastructure standard for automated regulatory compliance under the EU AI Act. We leverage the high-performance Solana network and decentralized IPFS storage to create an immutable audit trail for AI model training and deployment.

Our solution transforms the static legal requirements of Annex IV into a dynamic Compliance-as-Code pipeline. The system intercepts telemetry directly during neural network training (Loss/Bias), autonomously analyzes it via a multi-agent RAG architecture (Llama 3.3 70B), issues a safety verdict (Trust Score), and anchors the status on-chain.

This eliminates the risk of human error, protects businesses from existential fines of up to €35M, and allows compliance officers to generate a legally admissible PDF certificate for EU regulators with a single click.

We didn't just build a concept; we shipped an End-to-End process:

1. 🧠 Live ML Integration: The adapter script (demo_ml_training.py) integrates directly into a real model's training pipeline (scikit-learn MLPClassifier) and streams bias metrics in real-time.
2. 🔗 100% Web3 Proof: No local mocks. All AI safety JSON reports are automatically hashed and uploaded to the global decentralized IPFS network via Pinata, returning a cryptographic CID.
3. 🛑 System Lockdown (Red Alert): The frontend dashboard reacts instantly to the AI judge's verdict. If the Trust Score < 80 (discrimination detected), the system triggers a visual lockdown and records the failure on-chain.
4. 📄 Legal PDF Generation: The bridge between Web3 and compliance lawyers. The built-in generator allows officers to download a ready-made PDF report containing all cryptographic proofs (IPFS CID and Solana Tx) for EU regulatory bodies.

1. 🧠 AI Component: Llama 3.3 (Groq) autonomously analyzes intercepted training telemetry (Loss/Bias) in real-time.
2. ⚖️ Decision Making: The Agentic RAG acts as a judge, independently issuing a "Trust Score" and making a definitive "Pass/Fail" compliance decision without human intervention.
3. ⚡ On-Chain Action: If the AI detects discrimination (Trust Score < 80), it autonomously triggers a transaction to the Solana program.
4. 🔒 State Change: The Solana smart contract updates its state, permanently recording the restriction and locking the dangerous model out of production.

Deploying the automated compliance system:

Step 1: Environment Setup

git clone https://github.com/Sulemenov/AnnexIV-Core.git
cd AnnexIV-Core

Create a .env file in the 2_backend_ai folder and add your keys:

GROQ_API_KEY=your_groq_key
API_KEY=your_pinata_api_key
SECRET_KEY=your_pinata_secret_key

Step 2: Start Local Blockchain & Smart Contract (WSL/Linux)

⚠️ Requires WSL2 on Windows or native Linux/macOS

cd 1_solana_program
solana-test-validator --reset --ledger ~/test-ledger
# In a new terminal tab:
solana program deploy target/deploy/annex_iv_registry.so

Step 3: Start AI Backend

cd 2_backend_ai
python -m venv venv
source venv/bin/activate  # Для Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

Step 4: Start Frontend Dashboard

cd 3_frontend_dashboard
npm install
npm run dev

Step 5: LIVE DEMO (Machine Learning) Open the dashboard (http://localhost:3000). In a new terminal, start training the credit scoring model:

python demo_ml_training.py

Watch the telemetry stream to the dashboard. Around epoch 15, the model will simulate "bias drift," triggering the AI agent to slash the Trust Score and initiate a System Lockdown.

This project directly implements the technical documentation requirements outlined in Annex IV of the EU AI Act:

• Point 2(b): RAG agents automatically document the logic of algorithms and key assumptions.
• Point 2(g): The system captures "validation and testing metrics" (Loss/Bias) dynamically during the .fit() execution, generating an immutable audit trail.

We believe in radical transparency. Current project status (Decentrathon 5.0):