Kofi E. Ampaabeng

Software & Systems Engineering Leadership

I architect and ship distributed systems — and build the teams that run them.

The “can-do” engineer-leader: I take on undefined problems in unfamiliar fields — often with the obvious tools off the table — learn them, and ship. 20+ years across regulated banking, data platforms, and applied ML.

Why me

Build and lead

Architect and ship distributed systems, and built three teams from scratch across two co-founded ventures and a Director role.

Governance is the architecture

I treat engineering standards the way a bank treats controls — ADRs, tickets, tests, append-only audit.

Production proof

A core-banking and regulatory platform live for banks across Ghana and South Africa — real clients, air-gapped deployments, Bank of Ghana submissions.

Distributed data at scale

Built RegAuth (Spark/Solr over the full U.S. federal legal corpus) and Vigil (live streaming NLP/LLM), mentoring junior teams to deliver both.

Range under constraint

Led a WebSphere→WildFly migration with no prior experience in either; built a custom queue when Kafka and ActiveMQ were off the table.

Architect to the numbers

Right-size cloud-native (Azure/AWS) vs. self-hosted to fit budget and business indicators — not dogma.

The moat

A PhD economist who hand-builds the pipeline: measurement rigor most engineers lack, engineering depth most economists lack.

Leading with AI

AI agents write much of my code. My job is the system around them: what they may do, what they must prove, and what only a person decides.

Agents in parallel, under written rules

My engineering runs through several AI agents at once, each on its own ticket. They work under rules I wrote, and the rules run automatically: an action that breaks one is stopped before it happens.

Controls before speed

Every change starts from a ticket and a reviewed design. Separate agents plan, write and run the tests, and the plan is approved before a test is written. A change merges when its tests and acceptance criteria pass; design approval, production deploys, credentials and spend stay with a person.

Tools I built to run it

Switchboard puts every open thread across the agents on one board: the decisions waiting on me first, each agent's task card, every open merge request, and a log of decisions. CW DevOps takes a ticket to a tested pull request that merges only through a compliance engine.

Models I train, not only call

In SYNCopated, a note listener for pitch detection, an onset network on Basic Pitch's design being retrained on 30 hours of licensed, rendered and augmented piano audio for on-device inference, and a model that proposes the music regions on a scanned page, fitted on bands people confirmed. In Vigil, a Studies framework: people label samples in the app, and each study trains and evaluates its own classifier on them and records every run, so the models keep learning from the people who use them.

Training data nobody had to play

For the note listener, an AI agent plays note scripts on a digital piano over a MIDI cable, started over Wi-Fi, so the recorded audio comes back labelled with every note it holds. That audio was key to the listener's accuracy: on two pianos it never trained on, it recovers 88 to 96% of notes across the registers, at no more than one false note in five minutes.

What it has shipped

More than ten systems since February 2026. Two have clients: Fund Manager, a credit-union management platform being deployed for one of Ghana's largest credit unions, and CWC Rekon, reconciliation across mobile money, payment gateways and core banking.

Selected work

All projects →