Past experience
WPP
Senior Vice President, Product, London · 2022 to 2026
Grew adoption of an eight product enterprise SaaS portfolio used across 70+ markets, with annual users up 60% and annual plans up 77%. Set strategy and roadmap for a portfolio run three ways: revenue products on a P&L, adoption products designed for uptake, and platforms that let other products speed up their own roadmaps and run more efficiently. Led 100+ people across product, engineering, data science, design and operations on three continents. Shipped a production LLM agent that translates audiences across nine ad platforms, a governed Data Catalogue open to users, applications and AI agents, and privacy first clean room activation. More than doubled data product revenue and led an AI assisted engineering programme that improved cost efficiency by 26%.
- +60% growth in annual users
- +77% growth in annual plans
- 100+ people, US, UK, Europe and India
- +100% data product revenue
- $4B+ annual managed media on the campaign platform
- $25M investment base across the portfolio
WPP Media
Vice President, Data Science, London · 2012 to 2021
Built a global data science organisation from nothing to around 100 people across 10 markets. Productionised reinforcement learning for real time investment allocation, delivered optimisation systems for a major global client, and built self service automation tools that generated tens of millions in incremental revenue.
- 1,200+ hackathon participants across 21 countries, 40 events over 10 years
- 0 to 100 people across 10 markets
- 7 figures annual savings for one client
Accenture
Senior Manager, London (promoted from Manager, 2010) · 2007 to 2012
Led digital, analytics and transformation programmes. Delivered price elasticity modelling, trade promotion optimisation, demand forecasting, media mix modelling, e-commerce optimisation and complexity reduction. Won and delivered multi million pound programmes, repeatedly expanding scope within existing clients.
- +1% gross profit increase from a 10% portfolio reduction
- 20,000+ food and non-food price and promotion models
- 20% investment efficiency & $10M incremental sales
Nectar and Loyalty Management UK
Customer Insight Manager, London (now Aimia) · 2005 to 2007
Led analytics team within the Nectar coalition loyalty programme and led engagements with partners across retail, energy, travel, automotive, financial services, telecoms and hospitality (e.g. Amex, BP, Hertz, Sainsbury's and Vodafone). Delivered acquisition, up-sell and churn prediction, segmentations and personalisation, for example triggered 1-2-1 communications linked to consumer lifecycle with a brand and built the business case that secured multi million pound investment in a new Insight and Communication business unit.
- 10+ coalition partners served
- £m+ investment secured for a new unit
T-Mobile
European Analytics Manager, London · 2002 to 2005
Led analytics team across 5 markets providing churn modelling and value segmentation to shape post/pre pay bundle and tariff pricing, and transformed the e-commerce channel. Device demand and portfolio models cut procurement costs.
- 300% uplift in digital sales
- $25M procurement savings
- 5 European markets
Aristocrat
Business Analyst, Sydney · 2001 to 2002
Implemented web based business systems and automated product intelligence that informed roadmap decisions.
Schlumberger
Product Manager, London · 1997 to 2000
Built and launched data led products using patented technology to improve subsurface imaging accuracy for oilfield exploration. Surveys produced data at a scale that needed supercomputers to process in a timely manner.
- 5TB per typical 4 week survey
- 2,000 acoustic sensors
- 4ms between recorded events
NIWA (New Zealand's National Institute of Water and Atmospheric Research)
Solutions Engineer, Wellington · 1996
Engineered New Zealand's first coupled weather and water catchment model at the national water and atmospheric research institute. A major energy provider deployed it to optimise hydro dam operations.
- 1st coupled weather and catchment model in NZ
Antarctica New Zealand
Antarctic field researcher, Wright Valley (MSc Physics, University of Auckland) · 1993 to 1995
Deployed a network of ten automatic weather stations across the Wright Valley at 77.5 degrees south. Analysed temperature, humidity, wind, pressure and radiation records to explain the winds that keep the valley dry, tracked a micro scale front moving through the valley, and set up a mesoscale model (RAMS) to compare against the observations. Presented at the AMS Conference on Mountain Meteorology, 1995.
- 10 automatic weather stations
- 2.3 m/s down valley front speed derived from moisture jumps
- 125,000 grid points in the valley model
AI builds
Built to practise what I advise: governance, evaluation and trust are what make AI adoptable.
AI Cost vs Accuracy Optimiser
2026
A harness that runs assessments of apps and their agent’s AI usage across a range of models measuring accuracy, tokens, cost and latency. Each app keeps its own test cases and scorer; the harness pins the judge model, checks every answer came from the model asked for, and prices tokens from a versioned table. The results page plots each app's efficient frontier and compares it with the setting the app runs in production, allowing at-a-glance identification of a more optimal model selection. By way of example assessments are shown for 3 apps across a mix of Claude and Gemini models
- +10 points Data Analyst on Claude Sonnet 5.5 low against its production model: 98% vs 88% of 52 questions, at 5% lower cost
Try it: https://modelcost.paulmartin.ai/
AI Quality Scorecard
2026
Grades any AI agent's answers against a gold set: how often it's right, whether every number it states comes from what it actually read, whether its citations hold up, whether it declines what it can't answer, and how its answers drift from run to run. It reads the evaluation runs my apps already save, so 99 runs and about 1,750 answers were graded without changing an app, and a blind hand check of 30 answers tested the grader itself.
- 18 changes between four runs of the same model on the same 52 questions: 9 answers got better, 9 got worse
- 97% agreement between the scorecard's generic grounding check and the Data Analyst's own, across 692 answers
Try it: https://evals.paulmartin.ai/
Red Team Evaluator
2026
Agents that attack my other AI apps so I can harden them. A fixed, versioned suite of probes from the OWASP Top 10 for LLM applications, including instructions hidden in the search results an agent reads. An Auditor Agent judges each reply against the app's own rules without knowing what the attack was, a Severity Agent judges what it would cost in production, and marker words hidden in each app's instructions give an objective check on both. Findings are triaged, flagged for my review, and a fix is re-tested on exactly the same probes.
- 3 of 3 → 0 of 3 runs where one attack got through on the Data Analyst, before and after a fix
- 22 of 28 verdicts where the AI judge agreed with my own blind read of the transcript
Try it: https://redteam.paulmartin.ai/
AI Control Tower
2026
One dashboard for all of my live AI apps instead of separate consoles. Each app now writes one log line per AI request (model, tokens, cost, time and outcome, never the question or the answer), collected into BigQuery alongside the request logs. The dashboard joins that with each app's live settings and health and with the other AI Ops apps' results: accuracy from the AI Quality Scorecard, cheaper settings from the Cost Optimiser, re-test advice from the Red Team Evaluator. Fixed alert rules flag what needs attention. Visitors see a made-up example; the live figures need an access code.
- 8 alert rules, from daily spend and failed requests to a recommended security re-test
Try it: https://controltower.paulmartin.ai/
UK Nutrition Evidence Assistant
2026
A retrieval-augmented (RAG) assistant that answers questions about UK nutrition — food composition, dietary recommendations and population intakes — grounded in official sources (CoFID, SACN, NDNS) with citations. Structured numeric questions run as exact database lookups rather than retrieval, keeping data markers such as trace amounts distinct from zero values.
Try it: https://uknutrition.paulmartin.ai/
UK Nutrition Data Analyst
2026
A natural-language analyst for UK food composition data. Ask a question in plain English and an LLM writes a SQL query, runs it against CoFID 2021, and explains the result alongside the query itself, the result table and an automatic chart, so every answer can be checked.
Try it: https://nutritiondata.paulmartin.ai/
UK Nutrition Research Agent
2026
A research agent for UK nutrition questions that need more than one lookup. It plans, calls two other agents via MCP (the Evidence Assistant for SACN and NDNS evidence, the Data Analyst for SQL over CoFID) and writes a short report.
- ~35s and about $0.17 per research run
Try it: https://nutritionagent.paulmartin.ai/
Flood Risk Checker
2026
Four agents assess flood conditions on English rivers based upon a user-specified postcode. The Meteorology Agent reads weather-model forecasts, the Hydrology Agent reads river and rain gauges (and never sees the forecast), and the Risk Agent tests Hydrology against gauge records and flood history. An Orchestrator Agent sees only their three verdicts and explains where they agree, where they genuinely conflict, and which view should carry more weight. 3 replays are available for previous real floods on the data as it stood at the time, to reveal what happened.
- ~90s and about $0.12 per assessment: four agents on one river
Try it: https://catchmentagent.paulmartin.ai/
Field Naturalist Agent
2026
Three agents identify flora & fauna from real iNaturalist sightings awaiting identification. Vision Agent sees only the photo and proposes species with the features they rest on. Range & Season Agent sees only those candidates and the sighting's place and date, and checks how often each has actually been recorded there at that time of year, and which look-alikes are common nearby. An Orchestrator Agent sees only the two verdicts, decides between a confirmed identification, a locally common look-alike or a genuinely notable record, and writes the result to the public catalogue.
- ~1 min and about $0.06 per identification
Try it: https://fieldnaturalist.paulmartin.ai/
Stranded Nights Simulator
2026
Uses decision science to identify optimal price points for travel bookings of “stranded nights”: the short unsold gaps between stays that no break on sale can fill. A synthetic estate of 38 holiday parks and 41,000 holiday homes is simulated night by night under today's pricing and five pricing interventions. Each is tested as it would be for real, on some parks with others held back as a control, and the test is replayed hundreds of times in a Monte Carlo to show how sure you can be. Worked examples price real gaps four weeks out. The walkthrough and 25 ready-made tests are open to everyone; live runs are for invited guests.
- 25 ready-made tests: every cluster of parks with every pricing intervention, 300 Monte Carlo replications each
- +0.15% true yearly gain from gap-aware pricing on the default settings: real, but too small for a handful of parks to confirm
Try it: https://travelbookingsim.paulmartin.ai/
Supply Risk Early Warning
2026
Most commodity tools show a price going up or down. This one looks at what sits behind it: where cocoa, palm oil, wheat and sugar are grown, what's happening in those regions, and which shipping chokepoints they pass through, to warn before a shock reaches the price. Three approaches (watching the price, concentration and the growing calendar, and live supply and route risk) are compared in a Monte Carlo backtest over 500 thirty-year synthetic histories. The same data is served as tools to a Claude agent over MCP, and every figure in its answers is traced back to a tool result. Synthetic data throughout; the tabs are open to everyone, and asking the agent live is for invited guests.
- ~65% vs ~25% of price shocks warned of at least 2 weeks ahead: live supply and route risk against watching the price alone
- ~4 wks typical warning before a shock; about $0.04 per question to the agent
Try it: https://supplyrisk.paulmartin.ai/
Brand Refresh Studio
2026
A pitch tool for a brand designer. It captures a prospect's website, then two agents read it independently: Visual Identity Agent sees only screenshots with the words greyed out, and Verbal Agent sees only the copy. A Strategist Agent names where the brand's look and its stated values agree and where they genuinely conflict: the opening line for a pitch. Package Agents then build the refresh one package at a time (palette and type, then a new logo, then a new homepage), each keeping everything from the one before, rendered as real pages with exact colours and fonts, plus a presentation for the meeting. For invited guests only.
- ~$0.16 per prospect: about $0.09 to read the site and $0.07 for a full set of package mockups
- 3 packages mocked up cumulatively, each keeping the last one's palette, type and logo exactly
Try it: https://brandrefresh.paulmartin.ai/
This Portfolio
Built as a data catalogue · 2026
Every tile on this site is a record. The page reads the records and draws them, so adding a role or a project means adding a row, and search runs across every field.