A PDPA-aware HR platform for Malaysian SMEs, packaged as single-tenant SaaS. Laravel 12 and Filament for the HR admin panel, a React employee web app, a Capacitor-wrapped mobile app for iOS and Android, Postgres underneath.
It is not a demo. A real company pilots it for real payroll, leave and attendance, hosted the way described here.
The build, in numbers
| Total commits | 145 |
| AI co-authored | 104 |
| Share of commits | ~70% |
| Scaffold to pilot | 9 weeks |
Every one of those AI-authored commits went through the same gate a human-written commit would: review before merge.
The repo also carries design documents for individual features, role permissions, attendance geofencing, multi-company support, each written before the corresponding code. That is the same discipline as writing the plan before the code, applied project-wide rather than to a single module map.
What "AI agent built" looks like day to day
Review. A gap-finder audit compared the API against the admin panel and found a real payroll parity bug. A human reverted the one agent-proposed fix that was technically correct and organisationally wrong.
Hardening. An authorization sweep caught managers who could approve claims outside their own team, before any real money moved.
Maintenance. When a seeder bug broke CI silently, the fix became a standing entry in the repo's own agent context, so the next session cannot reintroduce it.
Notice what those three have in common. None of them is about generating code. All of them are about catching what generated code gets wrong, and then making the catch permanent.