Skip to content
Availability: openOn-site / Remote

George M'sapenda — Software Engineer

I build web and mobile systems end to end, and I lead the teams that ship them. Most of my work sits where product decisions meet architectural ones: payments, messaging, and the retrieval layer behind AI features.

Get in touch
  • System DesignConstraints → trade-offs → decisions
  • AI Integra—tionRetrieval · prompts · evals
  • RAG & RetrievalHybrid search · ranking
  • Technical LeadershipArchitecture · standards · review
  • Payments & MessagingMobile money · SMPP · webhooks
01Selected work

Projects

Four systems. The case studies are on the work page — this is the scan.

All work

  1. BalloAds

    An advertising and messaging platform that speaks SMPP to a mobile network directly

    • .NET
    • C#
    • Entity Framework Core
    • PostgreSQL
    • Python
    • +4
    BalloAds interface
    In production2025 — presentLive siteCase study →

  2. Brutus

    One AI service for a whole product suite, behind ports rather than SDK calls

    • NestJS
    • TypeScript
    • Groq
    • Llama 3.1
    • PostgreSQL
    • +3
    ai · architectureBackend service
    In production2026 — presentLive siteCase study →

  3. Octo

    Event ticketing for venues where the network drops exactly when you need it

    • TypeScript
    • tRPC
    • Hono
    • Prisma
    • PostgreSQL
    • +3
    Octo interface
    In build2026 — presentLive site3 decision recordsCase study →

  4. Penda

    A money app whose AI assistant can act on your finances but cannot quietly rewrite them

    • Deno
    • Supabase Edge Functions
    • PostgreSQL
    • Row Level Security
    • Gemini
    • +2
    ai · architectureInterface in build
    In build2026 — present1 decision recordCase study →
02Practice

How I work

I build the system, I lead the people who ship it, and I write down why it is shaped the way it is. The three are the same job — a preference that never meets a constraint does not survive any of them.

01

Software engineering

Fast with AI, gated by review
I use AI aggressively for the drafting and keep every checkpoint that catches a bad idea. Generated code gets read line by line; if I can't explain it in review, it doesn't ship.
  • AI drafts, humans merge
  • Every generated line gets read
  • Faster to the first draft, same bar at the gate
02

Leadership

Lead from inside the work
Titles are easier to respect when the person holding one is in the same repo — I take real tickets and sit in the same review queue. The rest of the job is stating the goal, the constraints and the bar out loud, then carrying context across the gap: what the business needs into the backlog, what the constraint costs back into the room where budgets get set.
  • Real tickets, same review queue
  • The goal, the constraint and the bar, in writing
  • Trade-offs priced in money and risk
03

Architecture

Show the options you rejected
A design is only settled once the alternatives are written down next to it. I price the running cost, the failure behaviour and the lock-in while I am choosing — not at the first invoice or the first outage.
  • Every choice ships with its rejected alternative
  • Cost, failure and lock-in priced while choosing
  • The record has to outlast the person who had it
03Ask this site

Working, not claimed

Ask a question about my work and get an answer drawn from what is actually written here, with the passages it used shown alongside. If the answer is not on the site, it says so instead of inventing one — which is the hard part, and the reason this is a demo rather than a claim.

Try asking

AI in production

Running now

Updated 15 Aug 2026

What it reads
Everything written on this site — 179 passages — plus my CV, certifications, LinkedIn history and public GitHub.
What it answers
Questions about my experience, the projects, and the decisions behind them.
When it doesn't know
It says so. It will not fill the gap with something plausible.
How to check it
Every answer shows the passages it came from, so you can read the source yourself.

For the engineers: hybrid keyword and vector search over voyage-3.5-lite embeddings, fused with RRF, streamed over SSE. The full trace is on /sandbox.

04Tooling

What I work with

Grouped by the layer each tool sits in, which says a little more than an alphabetical wall of logos.

Interface & Mobile

Web and mobile from the same design language, including apps that have to keep working when the connection does not.

  • TypeScript
  • React
  • Next.js
  • React Native
  • Flutter
  • Tailwind
  • Offline-first

Services & APIs

Endpoint and payload design, and keeping the contract between client and server something the compiler can check.

  • FastAPI
  • NestJS
  • .NET
  • Node.js
  • Ruby on Rails
  • tRPC
  • REST

AI & Retrieval

Most of the work in an LLM feature is getting the right context in front of the model and being able to tell whether it helped.

  • Anthropic / Claude
  • Groq
  • OpenAI
  • Gemini
  • Vertex AI
  • RAG
  • Hybrid retrieval
  • Embeddings
  • Eval harnesses

Data & Infrastructure

Schemas shaped around how they will actually be read, and enough infrastructure to deploy them repeatably.

  • PostgreSQL
  • Supabase
  • Prisma
  • Firebase
  • Docker
  • GCP
  • Vercel

Payments & Messaging

Mobile money and SMS are the default rails in the markets I build for, which makes them an architectural constraint rather than an integration detail.

  • DPO Pay
  • Mobile Money
  • SMPP
  • SMS gateways
  • Firebase Cloud Messaging
  • Webhook reconciliation

Architecture & Practice

The habits that let a small team keep changing a system quickly without losing track of why it is shaped the way it is.

  • Hexagonal architecture
  • Event-driven systems
  • Monorepos
  • ADRs
  • CI/CD
  • Vitest
  • RSpec
  • TypeScript strict
05Writing

Playbooks and articles

Standards I would bring to a team adopting AI coding tools, and articles on things I have had to work out.

All writing