Provenance AI

AI Decision Agent with Tamper-Evident Verification

Comment

Inspiration

As autonomous AI agents are entrusted with high-stakes enterprise decisions — from multi-million dollar vendor procurement tenders to credit underwriting and medical diagnostics — a critical trust crisis has emerged: AI operates as an unverified black box.

When an AI agent chooses Supplier A over Supplier B or denies a loan, enterprises face severe liability:

  1. The Admin Tamper Threat: Traditional audit logs stored in centralized SQL databases or AWS S3 buckets can be silently edited, deleted, or falsified by root administrators, disgruntled employees, or hackers. Centralized logs offer zero legal non-repudiation in court or regulatory audits.
  2. The Privacy Dilemma: Writing raw enterprise datasets or personal sensitive data (PII, HIPAA records) directly onto public blockchains violates global privacy laws like GDPR and banking secrecy regulations.

We were inspired to build Provenance AI to solve this exact dilemma. Our tagline defines our mission:

"AI can make a decision. Provenance AI makes that decision verifiable."

What it does

Provenance AI is a cryptographic memory layer and tamper-evident audit ledger engineered specifically for autonomous AI agents. It bridges cutting-edge LLM decision-making with enterprise-grade EVM blockchain auditability.

Key Features & Workflow:

  • Multi-Agent AI Decision Engine: Utilizes Google Gemini 3.6 Flash to ingest multi-vendor bids, loan applications, or medical data, generating normalized scoring, weighted evaluations, and clear decision rationales.
  • Cryptographic Merkle Tree Engine: Hashes raw input datasets, origin evidence, and AI reasoning chains into SHA-256 digests, packaging them into a zero-knowledge Merkle Tree.
  • On-Chain EVM Smart Contract Notarization: Seals the 32-byte Merkle Root onto the Ethereum/Sepolia testnet via our deployed ProvenanceLedger.sol contract (0xC442ce42A6763e25664147b088DbD50B01C375e5). Raw datasets remain 100% private off-chain.
  • Real-Time On-Chain Gas Fee Estimator: Provides pre-execution transparency before signing transactions in MetaMask, calculating gas units, Gwei rates, and real-time ETH/USD costs.
  • Interactive Forensic Audit & Tamper Verifier: Allows auditors to simulate database tampering (e.g., retroactively altering vendor scores). The forensic diff engine re-computes Merkle proofs live, pinpointing the exact corrupted field and flagging the breach instantly.

How we built it

We built Provenance AI using a full-stack, enterprise-grade Web3 architecture:

  • AI & Reasoning Core: Integrated Google Gemini 3.6 Flash via Express server endpoints to perform multi-criteria analytical decision processing across diverse domains (Procurement Tenders, Medical Diagnostics, Commercial Loan Underwriting).
  • Cryptographic Engine: Custom TypeScript implementation of SHA-256 hashing and Merkle Tree calculation routines to construct verifiable Merkle Roots for every decision state.
  • Smart Contract & Web3: Written in Solidity (v0.8.20), compiled via esbuild/solc, and deployed on the Ethereum Sepolia Testnet. Connected client-side using ethers.js v6 with native MetaMask browser wallet support.
  • Gas Fee Estimation Utility: Built a live EVM RPC gas estimator (estimateGasForRecord) querying Sepolia fee data to compute exact gas unit requirements and USD estimates.
  • Frontend & UI/UX: Built with React 18, Vite, TypeScript, and Tailwind CSS, featuring high-contrast dark mode visualization, interactive Merkle Tree visualizers, and an integrated pitch deck & smart contract viewer.

Challenges we ran into

  1. Balancing On-Chain Immutability with GDPR/HIPAA Privacy:
    Challenge: Storing raw AI prompts or enterprise data on-chain is illegal under privacy laws, but off-chain logs can be manipulated.
    Solution: We designed a dual-layer architecture where raw confidential data resides off-chain, and only the 32-byte Merkle Root is stored on-chain.
  2. EVM Contract Type Conversions for Merkle Hashes:
    Challenge: Converting 64-character SHA-256 hex strings into EVM-native bytes32 types required precise padding and byte handling across ethers.js and Solidity interfaces to avoid gas estimation reverts.
    Solution: Standardized byte-formatting utilities (toBytes32) ensuring seamless ABI payload encoding.
  3. Accurate Real-Time Gas Estimation:
    Challenge: Public testnet RPC nodes can fluctuate wildly in gas pricing or timeout during contract simulation.
    Solution: Implemented a resilient fallback algorithm that queries live network fee data (getFeeData) while providing deterministic EVM gas unit profiles (~88.5k gas units) for smooth pre-execution previews.

Accomplishments that we're proud of

  • 🏆 Live Smart Contract Deployment: Successfully deployed and verified ProvenanceLedger.sol on Sepolia Testnet with full MetaMask signing integration.
  • ⚡ Instant Mathematical Verification: Built a real-time Forensic Tamper Verifier that detects even a single character change in a multi-gigabyte dataset within milliseconds.
  • ⛽ Pre-Execution Gas Fee Transparency: Integrated an intuitive gas fee widget giving users exact ETH and USD cost predictions before initiating Web3 transactions.
  • 🎨 Enterprise-Grade UI/UX: Designed a clean, dense dashboard with interactive Merkle visualizers, mathematical score explainers, and live PDF audit certificate generation.

What we learned

  • Why Centralized Databases Fail for AI Auditability: We realized that saving audit hashes in PostgreSQL or AWS S3 is insufficient for court-level non-repudiation because root admins can rewrite both logs and hashes retroactively. Public EVM consensus is mathematically mandatory for untamperable timestamping.
  • Merkle Proof Efficiency: Storing a single 32-byte Merkle Root on-chain can cryptographically anchor millions of data points and AI reasoning steps without incurring high gas costs.
  • Developer & Auditor Experience: Enterprise compliance teams require human-readable score explainers alongside raw cryptographic hashes to trust AI outcomes.

What's next for Provenance AI

  • 🌐 Multi-Chain Deployment: Expanding smart contract deployment to Arbitrum, Polygon zkEVM, and Base for sub-cent transaction costs and high-throughput enterprise batching.
  • 🔐 Zero-Knowledge Proofs (zk-SNARKs): Integrating ZK-SNARK circuit generation so AI agents can prove their reasoning complied with specific regulatory rules without revealing the scoring weights or algorithm logic.
  • 🤖 LangChain & LlamaIndex Plugin: Releasing provenance-ai-sdk as an open-source middleware plugin for popular AI agent frameworks (LangChain, AutoGen, CrewAI), allowing developers to enable on-chain notarization with a single line of code.
Share this project:

Updates