Production AI, computer vision, and LLM agent systems for the factory floor, the trading room, and the operations centre. Twelve years of shipping software; the past six in enterprise AI.
Twelve years of building, from industrial AI and computer-vision platforms to AI/simulation R&D, exchange infrastructure, verifiable inference, and 3D engines. Filter by domain, click for the full story.
Talks, writing, and field notes on cryptography, ML systems, and verifiable computing.
Two weeks at Hi Tech! Stuttgart Match Arena pitching Ciro AI to German manufacturers. Almost nobody asked about the LLMs or the vision models. Everyone asked about ISA-101 compliance, SAP integration, and whether the dashboards would still work on a Windows machine from 2013. Notes from the conversations.
Read the field notes →
Qwen2.5-14B has 281 transformer layers and roughly 14 billion learned weights. We proved a full forward pass on Starknet Sepolia using Obelyzk — chunked weight binding, recursive STARK aggregation, and a Cairo verifier compact enough to fit on L2. Here's how, and what we'd do differently next time.
Read the post-mortem →
Most explanations of GKR sumcheck stop at the algebra. This talk walks through the engineering side: how to wire it onto StarkWare's STWO proving system over the M31 field, why streaming MLE evaluation makes billion-parameter verification tractable, and the specific CUDA choices that cut prover time by 18×.
Read the notes →
The M31 field reduces modulo 2³¹−1, which means every multiplication fits in a single 64-bit register — no carry chains, no Montgomery tricks. We wrote ~20,000 lines of CUDA to exploit that: memory-resident traces, warp-coalesced Poseidon Merkle commits, and a sumcheck round kernel that runs at 92% of theoretical HBM bandwidth on H100.
Read the deep dive →
A 36-hour build: an autonomous treasury agent that signs trades only when it can prove the inference that produced the decision. Proof is generated by Obelyzk, verified on Starknet, and settled into a smart-contract vault. A small but concrete answer to the "can we trust the LLM with the keys?" problem.
Read the write-up →I'm interested in AI that earns its place on a factory floor, machines that know what they don't know, the slow careful work of weaving intelligence into systems that have been running since before AI was the word for it; or in the rare evenings when none of it feels urgent and I get to be home.
Enterprise AI Engineer
Enterprise AI is the work I keep coming back to. Building intelligence into the systems where the stakes are real — factories, exchanges, treasury operations, operations centres — and where the line lead has been doing the job for thirty years. The math underneath matters: rigorous types, GPU kernels, the right model for the right place. But the question I actually care about is whether the operator on the night shift can use it without a manual.
Twelve years shipping production software across AI and simulation R&D labs (Vertex Studio), industrial manufacturing (Ciro AI), cryptocurrency exchanges (Kraken), and early-stage on-chain protocols. Expert in production ML systems, GPU-backed inference, computer vision for manufacturing, multi-agent and LLM architectures, real-time analytics, and the unglamorous data engineering that makes any of it run on customer infrastructure.
Nothing is harder than building simple things. First principles. Design with types. Test before you trust. Reach for the smallest tool that fits. Most ML papers should be a two-hundred-line script before they're a system. Most enterprise deployments should be a single well-instrumented service before they're a platform. Love the rawness of CUDA and the pragmatism of TypeScript — and I'll die on a hill for Rust.
Splitting time between Ciro AI (production AI for industrial manufacturing) and Obelyzk (verifiable ML inference). Open to senior IC or technical-lead roles at small, ambitious teams building real-world AI.
Tools I work with daily, the ones I trust under load, and the opinions I won't relinquish.
Production ML systems for enterprise customers — from training pipelines to GPU-backed real-time inference.
Multi-agent systems and LLM applications for industrial workflows, NL-to-SQL, and decision support.
Industrial vision for quality control, defect detection, and operational telemetry on the factory floor.
Functional-first, design with types, develop with tests — and I'll die on a hill for Rust.
The full enterprise stack — from operator dashboards down to the storage layer.
High-availability AI infrastructure across cloud, edge, and on-prem manufacturing networks.
Also working with: Unreal Engine + Unity for synthetic-data generation, Substrate / Cairo / Starknet for the verifiable-inference work at Obelyzk, and Circle STARKs over M31 for ML proof systems. Domain-specific tools that I reach for when the problem genuinely demands them.