22. Building MOOUSER — verifiable AI inference with zero-knowledge proofs. Making it impossible to fake an AI decision.
Cryptographic proof that AI models produce correct outputs. Banks run ML models for credit, fraud, and AML decisions. Regulators ask "prove it." We turn trust into math.
Built ANS entropy codec for neural network weight compression — 17% lossless improvement over Brotli. Achieved best SP1024 result (1.0996 BPB) in OpenAI's model compression competition, beating the merged leaderboard.
I built MOOUSER from scratch as a solo developer — the ZK prover, circuit compiler, on-chain Solidity verifiers, Stylus WASM verifier (24KB), REST APIs, 4 SDKs, CLI tool, and full production infrastructure.
The system generates zero-knowledge proofs that an ML model (XGBoost, LightGBM, Random Forest, Logistic Regression, Neural Networks) produced a specific output on specific inputs. Verification takes 629ms off-chain or 3-6M gas on-chain.
The architecture uses a TEE happy path ($0.0001/inference) with ZK dispute fallback — you only generate expensive proofs when someone challenges a result. This makes cryptographic accountability economically viable.
Open source. Apache-2.0. Cryptographic infrastructure that banks depend on needs to be auditable.
EU AI Act enforcement begins 2026. US regulators are increasing scrutiny on ML-driven lending. Banks need to prove their AI makes correct decisions. Today the proof is log files. Log files can be edited. Cryptographic proofs cannot.