Three years catching financial crime at Postbank — South Africa's state-owned bank, 13M+ customers — through its publicly reported R109M cyber-attack loss period. Since 2023: building production AI systems, and the evaluation frameworks that prove whether they actually work.
Most AI people have never worked a real fraud case. Most fraud people can't build or test AI systems. I do both — and the overlap is exactly where the industry's hardest problem sits.
I know how real financial-crime investigations work — and I know how to design the tasks that expose where AI models fail at them. Overconfident causal claims, hallucinated citations, missing chain-of-custody discipline: I build the tests that catch it.
Forensically catalogued tens of thousands of transactions — terminal IDs, beneficiaries, geolocation — into the evidence chains behind syndicate prosecutions, serving as sworn deponent for SAPS, the NPA and the Hawks.
Detection architected from raw, fragmented data with SQL, Python, advanced Excel/Power Query, Power BI and Graylog — real-time monitoring stood up in an environment with no enterprise tooling available.
Local-first assistant routing across models by deterministic rules with an escalation ladder and mechanical validators. 200 of 200 scripted reliability checks pass — seeded, so anyone can re-run the whole suite with one command.
Screening service for OFAC SDN and FIC targeted financial sanctions with fuzzy identity matching, built against a 336-test suite — compliance logic treated with the same evidence discipline as a fraud case.
This work is personal. I was the primary fraud analyst for a national institution serving millions of pensioners, beneficiaries and families. A systemic vulnerability is never just a technical flaw — it is a direct threat to the people least able to absorb the loss. My conviction is simple: the cost of fraud must fall on the perpetrator, never the victim. I hold AI systems to the same standard I held fraud controls — prove it works, or assume it doesn't.
I'm available for fraud-investigation expert work, AI-model evaluation, and remote roles on the seam between the two.
