Hi, I'm Paul

I'm an engineering manager who still can't keep his hands off the code. I've worked across Python, Java, AWS and the Frappe framework, and the job has always been the same: find the problem that's actually worth solving, then build the fix. That's what engineers are for. Lately, the problem I can't put down is making AI genuinely useful inside business software.

What I build

My current answer is Frappe Assistant Core (FAC), an open-source app that connects assistants like Claude to Frappe and ERPNext. You can ask "who owes us money?" and get a real answer from your real data, instead of a confident guess. The documentation has everything you need to try it.

I also built the runtime behind FAC Cloud, the hosted version. It decides which model handles each question, keeps the costs sane, asks a human before the agent does anything risky and remembers what matters between conversations. That part is closed source, and FAC Cloud is offered commercially through Promantia. If you want it in your business, talk to them.

How I build now

A lot of the code I ship these days is written with AI agents. I plan the work, split it into tasks, hand those to agents and spend my time reviewing what comes back. It's a bit like managing a team that types at the speed of light, never gets tired and is sometimes completely, cheerfully wrong.

That last part is the job now. AI made writing code cheap. It did not make checking code cheap. The lesson that cost me the most: a passing test suite told me three features worked when none of them did, because the tests set up a situation that production never reaches. Now every fix has to make a test fail before it's allowed to make it pass.

The notes here are made the same way. I draft them with Claude from my own commit history, then check every claim and number against the code before anything goes up. If you spot something wrong, tell me. That's how the next note gets written.

Currently

• Building and maintaining two things: FAC Cloud, a SaaS product, and Frappe Assistant Core, an open-source app.
• Learning how to build, deploy and maintain LLM-powered software that holds up in production.
• Teaching agents to ask for permission instead of forgiveness.

What this space is

This is my journey, written down as I go: what I learned, what I got wrong and what I'd do differently. Expect more mistakes than victory laps. The mistakes are where the useful lessons are.