AI/ML Leader

AI/ML Leader

I build machine-learning products and lead the teams that build them. Over the past decade-plus that has spanned biomedical engineering, computer vision, e-commerce, and application security — different domains, one craft: turning algorithms, ML, and deep learning into things people actually use.

Today I’m VP of AI & ML at Cycode. This site is the canonical record of that work — browse the portfolio, categories, timeline, or topics.

How I think about the work

When I started, the craft was rule-based heuristics. It got renamed computer vision, then machine learning, then deep learning, and once foundation models arrived, simply AI. Each wave took half the previous craft with it — the equations that used to get drawn on whiteboards, the architectures assembled block by block, the fleets of narrow models trained one task at a time. None of it is worth mourning. Every shift handed over better tools for the thing that actually matters: building products that work.

What survived all of it is depth. Understanding how a technology genuinely works is what makes it possible to see the thing sitting just beyond its reach — or the move that is unfashionable enough that nobody is attempting it, though it would probably succeed. Usually that move already exists as a solved pattern in a different field. That is where Peptriever came from: protein binding was only ever evaluated one pair at a time, which can never answer what else a protein might bind to. Borrowing the way search engines narrow millions of candidates down to a shortlist made binding search possible at scale — trained on a home gaming machine, with results competitive with the state-of-the-art binding models of the time.

Today the binding constraint on agents is not capability, it is reliability. A well-scoped agent does the right thing 99.9% of the time; grant it full autonomy at scale and the remaining 0.1% is what ends up in the news. That does not get fixed by waiting for a better model — nothing is ever quite good enough, and waiting is how you never ship. It gets designed around, so that being wrong stays cheap: irreversible actions gated behind a person, verdicts that lower priority instead of discarding information, output that lands in the review mechanisms teams already trust. That is how real automation ships today — and as reliability improves, those controls come off deliberately, one at a time, gaining speed in return.

What I’ve built

  • Cycode (VP of AI & ML) — I lead the AI & ML organization behind a suite of shipping AI agents: Maestro, the AI Exploitability Agent, a Remediation Agent, generic secret detection, and the AI Leak Analyser. One customer cut triage from over three days to under an hour and reduced MTTR for critical risks by 99.4%. Details and sources
  • Shopify — led the Product Understanding team: the ML foundation for Shop App and Shop AI, running categorization, attribute extraction, and embeddings on billions of products. Write-up
  • Donde Search — VP R&D, leading ML and engineering on ML-powered product discovery; the company was acquired by Shopify.
  • Patents and research — inventor on multiple granted patents across neural-network training, synthetic data, and product discovery; co-author of Peptriever (Bioinformatics 2024, 40(5), btae303).

Get in touch