Eighteen Years Shipping Regulated AI

A senior engineering leader I once worked with had a habit. When candidates told her they had ten years of experience, she would ask whether it was ten years or one year repeated ten times. The candidates who could answer the question well were the ones she hired.

This article is the answer to that question, written for executives evaluating senior AI hires. It is not a CV — a CV is a list. It is the eighteen-year arc that the list does not show: which environments produced which lessons, where the work changed the system, where the system changed me. If you are hiring at the head-of, director, or vice-president level, this is the kind of narrative you should be able to extract from any candidate before signing the offer.

The arc

### 2004–2010, Reserve Bank of India — Quantitative Risk and Supervision

Six years in central banking. Not as a consultant. As an officer in Quantitative Risk Management, Supervision Policy, Payments, Government Banking and Public Debt Management. The work taught me what regulators actually look at when they look at a model. They do not read the math; they read the documentation, the lineage, and the change history. If a model has clean inputs, clear assumptions, and a paper trail that survives a Friday afternoon with a sceptical supervisor, the math will be fine.

Two colleagues from that period — Shubhabrat Agrawal and Manoj Poddar, both General Managers at the Reserve Bank — wrote LinkedIn recommendations that I have kept on my profile since 2013 and 2017. They are not glowing testimonials. They are short, specific, and the kind of references regulated employers verify.

What this taught me: how a regulator thinks. This is the lens I have brought to every AI system I have built since.

### 2011–2018, ING / EY — Risk transformation in commercial banks and Big Four advisory

Seven years across two seats. At ING I led regulatory transition programmes — ECB, DNB, internal models. At EY I was Senior Manager in the Quantitative Advisory practice, leading multi-country change programmes for banking clients on Basel III/IV, IFRS9 and stress testing. The two seats taught complementary lessons.

The bank seat taught me what it costs to live with a decision for four quarters. The advisory seat taught me what it costs to recommend a decision someone else has to live with. Most consultants only learn one of these. Most internal leaders only learn the other. The combination is the entire reason a fractional model works at all — you are buying the seat that has done both.

### 2016–2017, Deutsche Bank — FRTB and IRRBB

Eighteen months in London leading the front-to-end workstream for FRTB and IRRBB regulatory capital at the investment bank. Designed both standardised and internal model approaches. This was capital — the part of the balance sheet the regulator cares most about because it is the buffer that protects depositors from the bank's worst day.

The lesson from Deutsche Bank was about scope. The regulation was specific. The implementation was specific. The deadlines were specific. The political environment around getting it done was anything but. Every senior position involves a translation problem between a precise technical reality and a non-precise political environment. The translation is the job.

### 2018–2021, Rabobank — IFRS9 Calculation Engine, €400B+ portfolio

Two and a half years as Product Owner of the IFRS9 Calculation Engine. The platform served 8,000+ employees and computed expected credit losses on more than €400B of loans. Agile delivery, ECB and DNB regulatory compliance, monthly close cycles that could not slip.

The lesson from Rabobank was about scale. A model that works on one portfolio at one bank in one quarter is a model. A platform that runs every month for years across every portfolio is a system. The transition from "model" to "system" is where most projects fail. Most candidates have shipped a model. Fewer have shipped a system.

### 2022–2025, Philips Personal Health — Enterprise GenAI Transformation Advisor

Three years on the front edge of the generative AI transition inside one of the largest medical device companies in the world. Founded the GenAI Champions Community: 200 members across business units, trained more than 500 employees on GenAI-powered workflows, produced approximately €500K in annual savings, achieved 80–99% metadata accuracy on the digital asset estate, reduced production cycles by roughly 35%, and improved compliance metrics by 45 percentage points.

The lesson from Philips was about adoption. The technology was not the bottleneck — Azure OpenAI and the surrounding stack worked. The bottleneck was always the same thing: getting people who already had a job to use a new tool that initially made the job harder before it made it easier. The Champions Community was the answer. Two hundred volunteers across functions who knew the new tools, knew the old workflows, and knew their colleagues. Top-down rollouts were less effective than peer-to-peer coaching at every measurable point in the programme.

### 2024–2025, ASN Bank and ING Bank — back into regulated environments

Five months at ASN Bank as Senior Model Validation Expert (ECB remediation, A-IRB credit risk models, validation standards). Six months at ING Bank as GenAI Platform Lead leading the enterprise deployment of Azure OpenAI and Microsoft Copilot Studio. Built more than 30 pre-built agents, achieved 65% training completion within six months across a 40-person team.

The lesson from these two engagements was about the gap between regulated and unregulated AI. Regulated environments — banks especially — already have the muscle for model validation, change control, audit trails. They do not have the muscle for the speed of generative AI. Unregulated environments have the speed and lack the muscle. Both ends need to meet in the middle, and the people who can broker that meeting are scarce.

### 2024–present, Amazon Ring — Strategic AI Transformation Advisor

Currently advising Ring on GenAI-driven content operations. Managing 2,500+ digital assets and the surrounding metadata, compliance and brand safety automation. The 50% efficiency improvement on content creation is measured against the pre-GenAI baseline. The work is ongoing.

The lesson here is recent enough that I am still drafting it. What I will say is that consumer-product GenAI at Amazon scale operates on a very different cost-and-speed envelope than regulated GenAI at a European bank. Both are real environments. Neither makes the other obsolete.

What the arc gives you

A candidate's arc tells you three things a CV cannot.

The first is what the candidate has done at scale. Not led, not advised — produced. €500K in annual savings is at scale. €400B in loan portfolio is at scale. Two hundred Champions is at scale. Anything below those orders of magnitude in a senior candidate is not at scale.

The second is the breadth of regulators the candidate has lived under. RBI, ECB, DNB and the EU AI Act each enforce in different ways. A senior AI hire who has only lived under one of them will be surprised by the others.

The third is the mix of seats — operator, advisor, founder, regulator-adjacent. The mix is what produces translation skill. Most candidates have one or two seats. Three or more is uncommon. Four is rare.

The patent and the paper

For executives who want a single artifact that compresses everything into one place: European Patent EP26162901.8 covers the TAMR+ methodology — temporal-adaptive multi-hop reasoning over regulatory knowledge graphs. The accompanying academic working paper is SSRN 6359818. The patent is the technical proof. The paper is the peer-reviewed framing. Together they are a checkable shorthand for "this candidate has produced original technical work in the exact domain you are hiring into."

You do not have to read either. You have to know they exist and verify them at the registry.

A note on what this is not

This is not a complete record. It leaves out the consulting, the open-source contributions, the founding of two AI companies, the EU residency, the languages, the public talks. It is the operator arc. The other arcs exist on the LinkedIn profile and in the formal CV.

If you are evaluating a senior AI hire and you cannot extract this kind of narrative from any candidate in a sixty-minute conversation, the candidate has not done the work. Hire the one who can.

Related articles