I design the models behind expensive decisions: who gets an insurance quote and at what price, which patient gets seen first, which household is a credit risk. I trained as an actuary first, so before I trust a prediction, I ask what happens if it's wrong. I've built this for insurers, a national hospital, and alongside the engineers at Vondetta.
I trained as an actuary, then taught myself to build the models an actuary would want to audit. One discipline asks what happens if a number is wrong and who pays for it. The other asks how to actually build that number, from raw data to a working system. I use both, on purpose, on the same problem, which is what makes me useful on a data science or ML team rather than just an insurance one.
Today, I build production software at iLabAfrica, Strathmore University, end to end, database through dashboard, for clients across government, healthcare, and conservation, and decide where the machine learning sits inside it. I'm also part of Vondetta, an AI research brand where a small group of Kenyan engineers builds homegrown, grassroots decision-support AI, for everyone from a mama mboga tracking her margins to a county official setting policy.
My rule: if a client can't follow why my system made a decision, I haven't finished the job.
The problems I was actually handed, roughly in the order they happened.
My career is still short on titles and long on finished work. In under three years I went from reading hospital claims for fraud, to pricing insurance applications, to building AI decision systems, to running my own company doing all three at once. Every stop asked a version of the same question: what does this number mean for the person who has to live with the decision it produces? That's the question I bring into any room.
What the problem was, what I built, and what the result actually means.
The problem: Getting an auto insurance quote in Kenya meant a broker manually reading your logbook and ID. Slow, and error-prone. What I built: A platform that reads documents automatically, searches a 105,000-document knowledge base for the right match, and prices it with four ML models.
The problem: Small African insurance agencies can't afford enterprise underwriting software, so one agent can only process a handful of applications a day. What I built: An underwriting engine trained on 58,592 real policies that points to similar past cases as evidence, the way an experienced underwriter explains their reasoning.
The problem: A €25.8 million insurance portfolio was losing customers, and nobody knew which channel, agents or brokers, was worth more over time. What I built: A platform predicting churn and claims, plus a lifetime-value model comparing both channels directly.
The problem: A shipping company needs a premium quote. Too low and the insurer eats future losses; too high and they lose the client. What I built: A full pricing model, plus a live dashboard where a non-technical stakeholder can adjust assumptions and see the price change instantly.
The problem: Trading strategies often look great on paper, then fall apart when real conditions change. What I built: A signal blending earnings quality with short-term price patterns, stress-tested against that exact failure mode.
The problem: Bitcoin swings wildly. Anyone pricing an option needs to know how long a shock takes to calm down. What I built: A volatility model over three years of daily data, comparing four modelling approaches to find the one that actually fit.
The problem: Kenya runs a massive national housing survey, but nobody had turned it into an early-warning tool. What I built: A vulnerability score covering 21,347 households across all 47 counties, built to be trustworthy enough for an insurer or lender to price against.
The problem: Community health workers see patients with no smartphones, no internet, and no shared language with a distant doctor. What I built: An AI triage assistant following WHO's child-illness protocol, working over basic USSD in English, Kiswahili, and Kinyarwanda.
The problem: Testing river water by hand is slow, so pollution can go undetected for months. What I built: A model reading satellite images to estimate water quality, no site visit needed, across 6,400 location-month readings.
The problem: A Kenyan county couldn't tell, mid-year, whether it was on track to hit revenue targets. The numbers only came together at year-end.
The problem: A perishable-goods transporter wanted to know if their manual route planning was any good, or if they should pay for logistics software. What I built: A mathematical proof comparing their method against the true optimal solution.
The toolkit, and what each part is actually for.
Where I explain my own work, so you don't have to take my word for it.
People who directly managed or worked alongside me. I don't publish their contact details here. Request one below and I'll be notified by email to share it with you directly.
I'll get an email letting me know you'd like to reach this reference, and I'll follow up with you directly.
If you're an insurer, a health system, a government office, or a team making a high-stakes decision with imperfect data, that's the work I want to hear about.