CV

AI evaluation scientist — LLM and agent evaluation · experimental design · Bayesian decision theory. Hong Kong permanent resident; UK and Israeli citizen; open to Hong Kong, Singapore, London, Tel Aviv or remote. [email protected] · LinkedIn · GitHub · Scholar. The two-page PDF is on the contact page.

Summary

Statistician (Warwick PhD, Oxford BA, 1,750+ citations) who designs AI workflows and measures whether they actually work, including when the result is bad news for my own system. At Booking.com I designed the evaluation for an agentic LLM recommender and used it to redesign the system: task completion up about 2.7×, latency down an order of magnitude. At HSBC I design and measure LLM workflows for compliance and fraud. My research is on evaluating and governing autonomous AI agents — leak-proof scoring, power analysis, checks proven against adversarial candidates, and criteria signed before the code exists — and is written up as A Court for Agents, in which every number can be recomputed from a public record.

Experience

Senior Data Scientist (contract), HSBC — Hong Kong, August 2026 to present. Commercial banking data and analytics. Designed and measured a retrieval-augmented workflow, with guardrails, that answers third-party vendor compliance questionnaires. Designing and measuring improved address remediation to catch fraud: raw customer address data turned into signals of financial crime.

Senior ML Scientist (contract), Booking.com — Tel Aviv, March to June 2026. Designed a multi-arm offline evaluation framework (LLM-as-judge scoring, variance decomposition, automated prompt optimisation) to compare configurations of an agentic LLM recommender and choose models. Used it to redesign the system from a multi-step agentic loop into a leaner deterministic workflow, from research prototype to production. Separated a genuine model regression from an evaluation artefact through ablations; defined the metrics and judge design for a new itinerary-extraction feature; built judge-based metrics, trace-level observability and same-day A/B harnesses.

Lead AI Engineer, Slice | Global Equity — Tel Aviv, July 2025 to January 2026. Regulated fintech. Designed the document-AI pipeline (OCR with Claude and Gemini extraction) that replaced manual onboarding: 500+ legal and financial documents in 20 minutes at 95%+ accuracy.

Machine Learning Engineer, Velotix — Ramat Gan, October 2024 to May 2025. Designed anomaly detection that surfaced unauthorised database permissions across a million-plus tables and 50,000-plus users, replacing manual access reviews at a major Singaporean enterprise.

Senior Data Scientist and team lead, Willis Towers Watson — remote, May 2022 to October 2024. Led the data science behind an LLM product that automates job evaluations against WTW’s grading methodology, deployed to enterprise clients. Early GPT-3 adopter; led a distributed team.

Earlier — Hong Kong and Tel Aviv, 2011 to 2022. Data Scientist, Entrée Capital (a deal-sourcing model found three pre-seed startups before announcement, all later funded); Business Analytics Senior Manager, HSBC; founder of Data Guru, a Hong Kong open-data company; Prudential Asia; EY; Postdoctoral Fellow, University of Hong Kong (causal inference, Mendelian randomisation).

Research

A Court for Agents, and jack (2026). A court for a self-modifying agent: a signed, append-only record; a 450-line checker any party can replay; every suggestion a forecast under a proper score; promotion between layers by anytime-valid betting tests on pre-registered criteria. The monograph recomputes every figure from the public record. The book · the founding essay.

assize (2026). The predecessor court for coding agents: every check proven able to fail against a planted fault (30 checks, 81 planted faults, all caught at the seal), a sealed judge measured on a host with a Bayesian power analysis (a half-achievement improvement in Crafter detectable at 33 episodes per arm), AI builder, reviewer and auditor sessions whose forecasts are scored by log-loss and whose models are elected by their record. The account.

Pre-registered evaluation of AI-built software (2026). Acceptance tests signed before any code exists, so an AI builder cannot fit the checks; a red team that found fifteen ways past my own verification gate, each turned into a seeded failure the gate must catch; a pre-committed adoption gate that refused to adopt my own tool. Signed Before the Code Existed · Fifteen Ways Past My Own Gate · The Gate That Said No.

Credence (2024 to 2026). Decision-theoretic tool selection for LLM agents. When a one-line greedy ablation beat the full agent under fair conditions, I withdrew the planned paper and published what survived: the value of exploration depends on attribution quality. Applied spin-offs: a governor for a coding agent’s tool calls that flagged an injected data exfiltration at 0.94 precision while interrupting 1.2% of safe sessions, and a model router at 96% lower cost than always using one premium model.

doorplate.dev (2026). Bayesian resolution of Hong Kong addresses, in English or Chinese, to the government gazetteer: 89.5% right and 3.8% wrong on real-world text.

Education and publications

PhD and MSc (Distinction) in Statistics, University of Warwick; BA in Mathematical Sciences, University of Oxford. 1,750+ citations, including BMC Medicine (2013) and the International Journal of Epidemiology (2013); the chain event graph papers in the Journal of Multivariate Analysis and Bayesian Analysis (2011). English (native), Hebrew (advanced), Cantonese and Mandarin (basic).

Skills

Evaluation design: offline evaluation frameworks, LLM-as-judge, variance decomposition, ablations, paired and A/B comparisons, power analysis, pre-registration, error analysis, proper scoring rules, ship/no-ship decisions. Statistics: Bayesian inference, causal inference, decision theory, value of information, sequential and anytime-valid testing. Tools: Python, SQL, R, Julia; Claude, GPT, Gemini; Arize, Langfuse.