From manual to AI-assisted engineering.
BMI Global Ed runs international student recruitment fairs: 60–80 a year across 18 countries, each bringing together hundreds to thousands of students and dozens of universities, with live payments. At most five offshore engineers look after the 14 systems behind them. A large backlog of technical debt slowed every change, from idea to live. I rebuilt how the team works around AI coding agents, and brought the team with me.
A small team, a big estate, and dates that don't move
Every fair runs on a set date, and the platform behind them takes live payments. The codebase carried a large backlog of technical debt, so with at most five engineers every change was slow to reach production.
Fixed dates
60–80 fairs a year across 18 countries, each with 600–4,000 student registrations and 25–100 exhibiting institutions.
A small team
At most five offshore engineers looking after 14 systems, from student registration to the exhibitor platform.
Technical debt
A large backlog of technical debt slowed every change on its way from idea to live.
A new way of building, not just a new tool
Engineers now start from a written specification and hand it to an AI coding agent, working alongside it rather than writing every line by hand. That meant changing the process and bringing the people with it, not just adding a tool.
A redesigned delivery process
I evaluated how the team worked, rebuilt the delivery process around AI coding agents, and chose the tooling to match.
Bringing the team with me
The team, based in the Philippines, was sceptical at first: using AI to write code was widely frowned upon there. The answer was spec-driven development and governed agents. Engineers write the spec, the agent builds from it, and its work passes the same checks as theirs.
Training that kept going
Two weeks of dedicated retraining got it started, and the training and process refinement carried on well beyond that.
Machine-written code, governed before it reaches live payments
The risk was never that AI would be slow. It was machine-written code reaching a platform that takes live payments with no rules and no trail. On the exhibitor platform, the coding agent works to the four-layer framework in FIG.01, on its own branches. Its changes pass the same automated checks as the engineers' own before they can merge, including an architecture validator that blocks any change breaking the module boundaries. On staging, release agents take over the repetitive release steps and flag risky changes, and a person in QA signs off anything going live. I architected the framework and the agents. The team built them.
What stayed hard
Governing the agents created a problem of its own.
- The specs started drifting. Once the specification became what the coding agent built from, agents were also writing the specs and tests, and those drifted from the code as the product moved. Left alone, that would have made the reference everyone relies on quietly wrong. So I stopped the authoring agents from touching application code, put every new spec through a structured interview and a separate challenge step, and set a regular triage that turns each mismatch into prioritised work. Drift still happens on a product that keeps moving; it is now found and fixed on a schedule instead of by accident.
Tech stack
Say hello.
Message sent.
Thanks for getting in touch.