Abdel · Applied AI & Verifiable Intelligence

Fifteen years building systems that can’t afford to fail. Now building the ones that can’t afford to lie.

Payments at bank scale. Ethereum consensus. Zero-knowledge proofs. Multiple open-source ecosystems grown from zero. A company-wide agentic transformation. Every chapter was about replacing trust with math. That work is now moving into AI safety, verifiable AI, world models, and physical AI.

Now

Head of Applied AI and Verifiable Intelligence at StarkWare.

Focus

World models. Verifiable computation. Agent systems you can audit.

Signature work

Co-author of EIP-1559. Originated Kakarot and Madara. Led StarkWare’s agentic transformation.

World models are the next substrate. Once AI starts acting in the physical world, integrity of execution stops being optional.

Operating thesis · 2026

01 · Direction

The question I keep coming back to. How do you verify what an autonomous system actually did?

Billions of autonomous agents are coming. They’ll operate in hospitals, on roads, inside financial systems. The current stack was never designed for software that acts on its own at that scale.

The hard problem isn’t capability. It’s integrity. Once a system acts in the world, the useful question stops being what it can do and starts being what it actually did. Did the audited model run? On the claimed inputs? Inside policy? Can its behavior be replayed and attributed after the fact?

Deployment is running ahead of the verification layer. That’s the gap I work on. World-model evaluation, agent harness engineering, and cryptographic proofs of execution for autonomous systems you can actually audit.

Three things have to come together to make this work. Systems discipline from shipping under real conditions. Cryptographic guarantees from the ground up. ML depth to actually implement the models. The intersection is narrow. It’s the work I’ve spent the last decade moving toward.

02 · Track record

The work has always lived in places where mistakes don’t stay theoretical for long.

Agentic transformation · StarkWare/ Now

Running the company-wide agentic transformation across engineering, BD, product, and marketing. Maturity framework, org-wide survey, workshops and training. Agentic changes merged into Stwo, the STARK prover securing Starknet in production.

Starknet ecosystem · StarkWare/ 2022

As Head of Ecosystem, originated Kakarot and Madara and grew both from whiteboard to independent teams with their own contributors and communities. Kakarot spun out as a company on a $5M seed. Ran the Exploration and Ecosystem teams in parallel, 16 people across the two.

EIP-1559/ 2021

Co-author with Vitalik Buterin. Led the Hyperledger Besu implementation. The fee-market reform has governed every Ethereum transaction since the London hard fork in August 2021.

Hellhound/ 2018

Co-founded inside ConsenSys R&D. A decentralized blind-computation platform. Run programs over homomorphically encrypted inputs on a network of nodes, with no one (not even the network operator) seeing the data. I designed and built the HHVM, the Paillier pipeline, the Kubernetes infrastructure, the on-chain proof contracts, and the Byzantine detection logic. First author of the Red Paper. Live demo at DevCon4 Prague.

Payments · Dejamobile/ 2012 · 2018

Architected a cloud-based NFC payment platform for French and European banks, conformant to Visa and MasterCard. PCI-DSS end-to-end, HSM integration, mobile obfuscation and white-box cryptography mitigations noted by the payment schemes during certification. Co-inventor of French patent FR 2 996 663 A1.

03 · AI safety

AI safety

Deployment is running ahead of the verification layer. That is the gap. Cryptographic proofs, governance tooling, and agent harness engineering for systems you can actually audit.

ProvableWorldModel/ rust

Commit-and-audit proof system for deterministic, quantized world-model inference. Verify what the model actually computed.

→

llm-provable-computer/ rust

Can you prove an LLM produced a specific output without running it again? Exploring verifiable inference via STARKs.

→

verifiable-intelligence/ rust

Verifiable open-weight LLM inference markets, built on the CommitLLM commit-and-audit protocol.

→

claude-md-compiler/ python

Compiles CLAUDE.md into a versioned policy lockfile. Enforces it against diffs, hooks, and CI. No LLM in the runtime path.

→