VeraNode

VeraNode solves AI supply chain vulnerabilities where unvalidated models from repositories enable backdoors, granting attackers infrastructure access. Traditional tools can't inspect serialized ML formats, and no detection exists for authenticity, risking deepfake fraud like the $25M Hong Kong CFO scam. Q1 2025 saw $200M+ in deepfake losses, with $40B projected by 2027. 67% of organizations fear AI model integrity issues.​

Enterprises struggle to verify GPT-4 vs. GPT-3 in APIs, prove training for compliance, ensure medical/financial liability, and protect IP in sharing. VeraNode provides decentralized ZKML proofs confirming exact model usage without exposing internals, preventing provider swaps to cheaper/tampered versions critical for finance, healthcare, and legal AI.​

In a medical example, hospitals register "Cancer Detection v2.1" hashes on Avail blockchain, generate zk-SNARK proofs per scan to verify outputs, pay $0.001 via x402, and store for audits—if tampered, proofs fail, with agents detecting anomalies. This counters "TrojAI" on Hugging Face evading scanners.​

The $250B AI model market by 2027 lacks verification competitors; Hugging Face offers docs only, MLflow metadata sans proofs, SageMaker no authenticity checks. VeraNode's edge: production ZKML, multi-chain via Avail, x402 micropayments, Lit IP protection, and compliance trails—overcoming ZKML barriers for enterprise trust

Built from scratch in a monorepo with frequent Git commits for ETHGlobal compliance, using Python/FastAPI for async backend, PostgreSQL/Redis/Celery for data/tasks. User Authentication handles JWT/Web3 logins and roles.​

Model Registration computes SHA256 commitments from uploads, stores metadata in PostgreSQL. Blockchain Module writes hashes to Avail contracts via Solidity, returning receipts for immutability—Avail's data availability boosted scalability.​

ZKML Proof Generator uses async Celery jobs with circom/zk-SNARKs to process model/input/output, storing proofs; GPU acceleration handled compute intensity. ZKML Verifier checks proofs against blockchain commitments cryptographically.

x402 Middleware verifies signed payments for endpoints, enabling agent micropayments. Lit Protocol encrypts weights with policy-enforced keys for IP safety. AI Agent Manager runs PyTorch adversarial tests, logging anomalies. Notification Service sends webhooks/emails for events; Analytics tracks verifications for dashboard APIs.

Integrated sponsor tech: Avail for chaining, Lit for encryption, x402 for payments—hacky custom ZKML circuits proved ML inference without full disclosure, piecing SDKs despite multi-protocol complexity (20% hackathon feasibility). Open-sourced on GitHub for verification