Head of Applied AI & Verifiable Intelligence — StarkWare
World models · AI safety · verifiable computation.
Closing the gap between what autonomous systems claim and what can be proven.
- Co-author and technical champion of EIP-1559 — the fee-market mechanism now securing Ethereum mainnet.
- Head of Applied AI & Verifiable Intelligence at StarkWare — applying STARK proofs to AI integrity, inference, and settlement.
- Former Ethereum protocol engineer (ConsenSys / PegaSys). 15+ years shipping production systems across payments, embedded security, cryptography, and blockchain.
- Originated Kakarot (zkEVM in Cairo) and Madara (Starknet sequencer) from zero; both now run with independent teams. Bootstrapped 15+ open-source projects and grew the Starknet developer ecosystem to thousands.
- Co-inventor, French patent FR 2 996 663 A1.
The question that drives the current work:
How do you verify what an autonomous system actually did?
Peer-reproducible reports and preprints. Full record on ORCID.
- Frozen-Backbone JEPA-Style Probes on HAM10000 (2026) — A nine-experiment ablation testing whether a latent predictor over a frozen vision encoder generalises on a leakage-controlled dermoscopy proxy; published with a full pretraining-contamination audit and an explicit "what is not yet claimed" boundary. →
10.5281/zenodo.20556968· code:derma-jepa - Toward High-Assurance AI: Safety by Design for Autonomous Systems (2026) — A framework for treating verification as a first-class system property rather than a post-hoc add-on. →
10.5281/zenodo.20512718 - Patch-Token Geometry of Four Released Vision Encoders (2026) — A multi-dataset artifact study comparing the patch-token representation geometry of DINOv2, I-JEPA, V-JEPA 2, and EUPE. →
10.5281/zenodo.20048509· code:latent-inspector - The Half-Life of Trust (2026) — Why durable AI verification needs mathematics-rooted, post-quantum foundations rather than hardware-rooted trust (TEEs). →
10.5281/zenodo.19682195
| Project | What it is |
|---|---|
| worldforge | Testable world-model workflows for physical-AI systems: providers, planning, evaluation, benchmarking, and diagnostics. |
| Lensemble | Federated, end-to-end JEPA world models trained across mutually-distrusting data silos — with a roadmap to STARK-based proof of each participant's contribution. |
| GenoLeWM | Action-conditioned JEPA world model for genomic edits: treat a DNA edit as an action over a frozen DNA encoder and predict the latent outcome. |
| CodeLeWM | Latent transition models over code edits — a scorer/reranker for candidate patches, not a generator. Published as infrastructure plus an honest negative result. |
| derma-jepa | Frozen-backbone JEPA-style probes on HAM10000: can a latent predictor distinguish stable lesions from changing ones under nuisance variation? Contamination-audited. |
| gpc_rs | Generative robot policies in Rust: diffusion policy + world model + evaluator. |
| asimov-sim-lab | Schema-backed inspection, validation, and deterministic evidence packaging for the Asimov v1 MuJoCo humanoid model. |
| mosaicmem | Geometry-aware spatial memory for camera-controlled video generation. |
| Project | What it is |
|---|---|
| jepa-rs | Rust implementation of JEPA primitives: I-JEPA, V-JEPA, C-JEPA, VICReg, EMA. |
| latent-inspector | Compare DINOv2, I-JEPA, V-JEPA 2, and EUPE representation geometry on the same image. Rust + ONNX. |
| latent-inspector-py | Patch-token geometry benchmarks for released vision encoders (DINOv2, I-JEPA, V-JEPA 2, EUPE). |
| attnres | Attention residual stream experiments inspired by Kimi / MoonshotAI. |
| turboquant | Rust implementation of Google's TurboQuant for KV-cache quantization. |
| nostrain | Coordinator-free distributed ML training over Nostr relays using DiLoCo. |
| jepa-notebooks | Interactive notebooks for JEPA architectures. |
Deployment is moving faster than the verification layer. That gap is the work.
| Project | What it is |
|---|---|
| llm-provable-computer | Verifiable LLM inference with STARKs. |
| awesome-ai-safety | Curated map of AI safety tooling: alignment, interpretability, red teaming, formal verification, zkML, governance. |
| eu-ai-act-toolkit | Open-source toolkit for EU AI Act compliance. |
| claude-md-compiler | Compiles CLAUDE.md into a versioned policy lockfile and enforces it against diffs, hooks, and CI. |
| Project | What it is |
|---|---|
| Kakarot | EVM interpreter in Cairo. Ethereum compatibility on Starknet via ZK proofs. |
| Madara | Starknet sequencer for sovereign appchains. |
| Raito | Bitcoin ZK client in Cairo. Verifies Bitcoin consensus inside a STARK proof. |
| Askeladd | Verifiable computation for Nostr Data Vending Machines via STARKs. |
| Cashu ZK Engine | Blind Diffie-Hellman key exchange in Cairo for Cashu ecash. |
| Project | What it is |
|---|---|
| collidervm | ColliderVM: stateful computation on Bitcoin without fraud proofs. |
| bitcoin-mcp | Bitcoin and Lightning MCP server. |
| nostringer-rs | Ring signatures for Nostr, in Rust. |
| nostr-mcp | Nostr MCP server. |
| bitcoin-honeybadger | Agent-based modeling for Bitcoin network resilience. |
| Project | What it is |
|---|---|
| parler | Multilingual voice intelligence on Mistral Voxtral — structured decision logs from French/English meetings. |
| Title | What it is |
|---|---|
| World Models, From Zero to Hero | A field guide to one of the most overloaded terms in AI. |
| The Half-Life of Trust | Why verifiable AI needs post-quantum foundations. |
| Math Is Humanity's Last Bastion Against Skynet | Why ZK proofs are the foundation for AI safety at scale. |
| Can LLMs Be Provable Computers? | Verifiable AI inference via STARKs. |
| Before Fighting Banks, Let's Understand How They Actually Work | A cypherpunk's guide to the financial system. |
| Time to Take the Nostr Pill | Why Nostr matters for freedom of speech. |
| Nostr DVMs Meet Verifiable Computation | STARKs for trustless Nostr services. |
| Cashu Meets STARKs | Zero-knowledge proofs for the Cashu protocol. |
Hellhound — applied cryptography / blind computation (2018)
I co-founded Hellhound inside ConsenSys R&D: a decentralized blind-computation platform for running programs over homomorphically encrypted inputs without exposing the data to the network.
I built the HHVM register-based bytecode VM, the Paillier homomorphic-encryption pipeline, the Kubernetes/GKE infrastructure, the Ethereum smart contracts for on-chain computation proofs, and the consensus logic for detecting malicious nodes. I was also first author of the Red Paper.
We shipped a live demo at DevCon4 Prague.
Close the gap between what systems claim and what can be verified.
Math scales. Goodwill doesn't.