Leading AI transformationat M-PESA, after twelve yearsbuilding AI products at Amazon and AWS.
Eight markets, 100M+ customers and $2B+ revenue across Vodacom FinTech Group, reporting to the CEO. From Alexa to Amazon Bio Discovery, an agentic AI application for drug discovery, with organizations of 100+ engineers, scientists and product managers. A master’s degree from Stanford Graduate School of Business, and published research from the Stanford Institute for Human-Centered AI on AI’s global economic impact.
RanLed the smart home features of the Alexa app, used by 50M+ people a month. Technical advisor and chief of staff to the Vice President of Amazon Smart Home and Health.
ResultMap View launched in the United States, November 2023.
RanFirst product lead of a computer vision and AI organization of 100+ scientists and engineers. Directed its $100M+ a year investment portfolio; managed 10 product managers.
ResultThree validation studies in npj Digital Medicine: Body and Movement in 2022, waist-to-hip in 2023.
RanOwned supply-chain strategy for a $700M+ a year Apple category across about 1,800 stores.
ResultIn-stock rate up 24% and market share up 3%. Allocated the iPad mini launch to each store, November 2012.
Patents and papers: six US patents, 2022 to 2026; three validation studies in npj Digital Medicine, 2022 and 2023; published research from the Stanford Institute for Human-Centered AI, 2025.
My approach
I lead AI transformation at two levels: setting the strategy with the CEO and executive team, and working hands-on in the forward-deployed engineer’s role, building the first systems with the teams who will run them.
Move AI from units of work to systems of work
AI on a unit of work, one step in one function, changes little while the handoffs around it stay the same. I redesign the system of work, the whole chain across functions, and measure it from the first request to the result.
Prime NowI took the stores’ spreadsheets, then automated a matched catalog, forecasts and a daily stock upload. Out of stock at partner stores fell from 25% to 5%.
Decide what “accurate” means before building
I start from the claim a customer will read, and choose the reference the model learns from and is judged against. When the labels are noisier than the claim allows, I change where the labels come from.
Halo BodyTraining labels moved from experts’ eyes to clinical X-ray scans. Published error: 2.2 points, half that of the best smart scale tested.
Put one customer measure above the model metrics
A model can beat its benchmarks while customers think the product does not work. I run one measure of the customer’s outcome for the whole system, and hold every model metric to it.
Alexa ShoppingI listened to 1,000 requests a week and wrote how each was judged. Success rose from about 50% to 85%.