I build products, not demos.

My County is published on Google Play and runs as a web app from the same Flutter codebase. It gives Kenyan counties, constituencies and wards their own local feed — posts, events, a marketplace, a business directory, attractions and real-time chat, all scoped to where you actually live. It is the project I have worked on longest, it has real users, and it spans four distinct pieces of engineering.

Support workers in NDIS and aged care finish a shift and still owe an hour of paperwork. ShiftNote AI takes the rough notes they actually write — "leon refused meds again. oats for breakfast. gp called." — and returns a structured, section-numbered progress note in the format their provider expects.

I built the whole product: Firebase email authentication with verification, a template engine so a provider can encode their own note format, per-account usage metering against daily and lifetime limits, and Stripe subscriptions with webhook-driven enforcement across a free and Pro tier.

The compliance side shaped the design as much as the AI did. Raw notes are minimised rather than retained, the output follows NDIS progress-note conventions, and the data model was built with Australian privacy expectations in mind.

A live platform with fifty registered students, which has generated over 130 courses to date. A student enters their institution, their course and what they already know, and it generates the entire syllabus from that: six to eight units, four to six lessons each, and an end-of-unit exam of ten to fifteen auto-graded questions on a countdown timer.

Generating a course is expensive if you do it naively, so it is deliberately cost-engineered — GPT-4o-mini rather than a frontier model, generation deferred until a lesson is actually opened, and content streamed to the page as it arrives instead of blocking on a complete response. A full course lands at roughly one to two dollars of inference.

Data sits in Supabase Postgres with Row Level Security on every table, so a student can only ever read their own progress, attempts and answers — enforced at the database rather than trusted to application code.

A bilingual AI assistant in English and Swahili, paid for the way people in Kenya actually pay: M-Pesa.

The payments work is the substance. It runs against Safaricom's Daraja API using STK Push — the server initiates the charge, the customer's phone prompts for a PIN, and an asynchronous callback later confirms or fails the transaction and credits the account. That means designing for a payment whose confirmation arrives out of band, minutes later, or never at all.

Around it: Firebase authentication and storage, document upload so users can ask questions about a PDF or spreadsheet, a credit ledger with usage analytics, and a real internationalisation layer rather than hardcoded strings.