I build systems that have to be right,
not just done.

I started in Aug 2017, freelance, building and maintaining web and backend systems for startups and small businesses across East Africa and abroad while I was finishing a computer science degree. Unglamorous and formative: cutting response times with caching and fewer database round trips, and migrating large systems between AWS and GCP in both directions. Nine years of engineering experience starts there, and almost all of it has been remote.

At Sonolabs I owned the computer-vision pipeline behind automated meal logging — segment each item on a plate, classify it by food type and portion — and curated and annotated a dataset of Kenyan foods, because no public one existed for the cuisine. At Naya Labs I was forward-deployed into client teams on three-to-six-month engagements across seven countries. On OmniArsenal, a US defence contractor's Go CI/CD platform, a defect that had been treated as a rendering problem for a month turned out to be a browser-storage key. Symptom and cause sat in different layers, which is the whole argument for monitoring you can read.

At PowWater the dispatcher picked the fill site, the driver and the route out of their own head. I automated that decision across fill-site assignment, route optimisation and fleet management, and delivery went from two to three hours to roughly 45 minutes. At Backbone Technologies I was one of two engineers on a US fintech platform, owning the backend, the AWS architecture and the payments stack: I selected Stripe and ran the migration when Chase discontinued WePay, held p95 API latency under 100 ms on an event-driven TypeScript backend, and cut monthly cloud spend by 30%. At DataQRL I am technical co-founder and CTO, part-time, and have designed and built eight modules of a financial-crime detection and compliance platform from an empty repository.

The turn came from a question the infrastructure could not answer. A biomedical knowledge graph hands you edges, and most of them assert co-occurrence rather than cause — a literature co-mention, a shared pathway, an arrow that is an artifact of column order. Point an estimator at one of those and it returns a plausible number with nothing behind it. I built NeoRx to identify drug targets over those graphs, which meant learning causal inference properly rather than around the edges: identification first, estimation second.

In September 2026 I audited it myself, and the stage that decides whether a causal claim is licensed at all did not hold. I withdrew every quantitative result that depended on it. The benchmark run behind those numbers had persisted nothing, so there was not even an artifact left to re-examine — which is why the replacement writes a checksummed record for every run. I rewrote the identification step, checked it against confounder, collider and mediator cases, and published it separately as CausalPhylo. The benchmark has since been re-run over the four diseases whose baseline comparison is methodologically sound; those figures, and exactly what they cover, are on /research. I would rather publish the correction than the number.