I design AI workflows and prove whether they work.
I’m a Bayesian statistician. For a team shipping an LLM or agent system, I build the evaluation that decides whether it ships, and when the answer is no I say so. For the question behind that one — how to oversee an agent that no human is watching and that is allowed to rewrite itself — I built jack, a court for a self-modifying agent, and I’m writing the book about it, A Court for Agents, in which every number can be recomputed from the court’s public record.
2.7×
task completion at Booking.com, after the evaluation I designed redesigned the system; latency down an order of magnitude. The work
1 paper
withdrawn when a one-line ablation beat my own agent, and the reason published instead. The Ablation That Beat the Agent
81 / 81
planted faults caught by the predecessor court's checks at its seal, and a half-achievement improvement detectable at 33 episodes per arm. The Court That Tried Itself
Who I work with
Teams shipping LLM and agent systems
You have a system that works in the demo. What is usually missing is the evaluation that says whether the next change made it better. I build that: metrics tied to the ship decision, an LLM-as-judge validated against human labels with its variance decomposed, paired comparisons with the power to mean something, ablations that tell a real regression from an artefact of the evaluation, and criteria fixed before the run.
Read: Signed Before the Code Existed · Make Your Agent Cheaper, and Measure It Yourself · The Ablation That Beat the Agent
Banks and insurers governing AI
The HKMA's ART framework (October 2026) puts accountability and human oversight first and is about to propose a registry of material AI agents and ways to intervene in them; MAS's guidelines take effect in October 2027 and make you answerable for third-party AI too. I build the evidence a second line can sign off: an audit trail of what an agent did and under which rule, the inventory and the obligations derived from it, change control for the rules themselves, and independent validation of vendor and in-house GenAI against pre-stated criteria. I've worked inside HSBC twice, and at Prudential and EY. Financial crime is where this started: address data turned into fraud signals.
Read: what that looks like · the audit trail (book, chapter 3) · change control for the rules (chapter 9)
Evaluation and oversight research
In the field's words: pre-registration made binding (acceptance tests signed before the code exists), mutation testing (every check must catch a planted fault), AI control (a judge the policy cannot reach), proper scoring rules on every forecast, and anytime-valid sequential tests for whether a layer has worked. jack runs all of it on a public record; assize measured what it cost in human attention; the book derives each piece from a result that was mostly settled decades ago.
Read: Every suggestion is a bet · Testing by betting · The evidence · Fifteen Necessities
Open to research and senior scientist roles in AI evaluation, oversight and assurance, and to contract and consulting work. Hong Kong, Singapore, London, Tel Aviv or remote.
If you need to know whether an AI system works, and to be able to show it — here is how I work, the two-page CV, and where to write.
Oxford mathematics · Warwick PhD in statistics, cited 1,750+ times · Postdoctoral fellow, University of Hong Kong · HSBC · Booking.com · Hong Kong permanent resident · UK and Israeli citizen
The book
A Court for Agents Law, scoring and the self-modifying machine
The monograph of jack, a court for a self-modifying agent, written as a construction: each chapter builds the smallest thing that answers what the last one left open, and every number carries the entry of the record it came from. Written with two Claude sessions under the court's own rules, every figure checked by script against jack's public record. A first draft, open to read while the layers cross.
First draft, November 2026. Next: chapters 6, 7 and 13 are completed from the C2a and C2b window exits; C2a ends on 15 November at the latest. The first edition is cut at C3's exit, when the agent arrives; it scores chapter 15's bets. Send a correction · Follow revisions
Read the draft →Start here
Five essays that carry the argument. Not the newest — the ones worth reading first.
Why We Stopped Using the Mathematics That Works
Start here literally. The mathematics for making decisions under uncertainty has existed since the 1960s and the field walked away from it — not after weighing it against something better, but because ImageNet made a different bet pay off and nobody looked back.
The Science Was Done. The Engineering Wasn't.
The argument. An agent allowed to change itself needs a core it cannot reach, an edge it may rewrite, and a score that is rigid everywhere — and nearly every piece of that was proved decades ago. The founding note for jack and the book.
Signed Before the Code Existed
The method. The acceptance tests were written, hashed and signed before a line of the implementation existed, then handed to a coding agent that could not change them. Pre-registration, made binding.
The Ablation That Beat the Agent
The integrity. A one-line ablation beat the agent I had announced a paper about, so the paper never went out and the record of why is here instead. What survived is narrower and better than the claim.
The Court That Tried Itself
The cost. The predecessor court reached its seal and a measured verdict, and it also cost 121 of my repairs in four days. The book begins where this essay ends.