Build the tools that build the software.

1 · Loop Engineering Platform

An AI agent writing code is only as good as the loop around it: give it a task, let it work, run the checks, feed the result back, repeat until the checks pass. Today every engineer wires that up by hand.

You'd build the platform that makes it reusable — a way to define a loop, run it safely, see every step it took, and measure which version works best.

2 · Slack Agent for Software Delivery

Requests arrive where people already are — in Slack. "The export button times out on large files." Today a human copies that into a tracker and it waits.

You'd build a Slack bot that turns that message into a real code change: understand the request, run it through the platform from Project 1, and report progress back in the thread.

The goal is to automate as much of that path as we safely can. So the bot triages by risk — low-risk changes it can verify and ship on its own, anything bigger stops at a human gate for review. Deciding where that line sits, and proving it's in the right place, is the most interesting part of the project.

3 · AI Scheduling Optimisation

Every service business answers the same question daily: who does which job, in what order? Get it wrong and technicians sit idle, or drive across the city twice, or miss a deadline. Today that's decided by hand, in a spreadsheet, by someone who's very good at guessing.

You'd build a scheduler that proposes the plan instead — balancing skills, location, travel time, priority and deadlines — and then prove it beats the manual version on real jobs.

The constraint that makes it interesting: it has to be general. The engine shouldn't know anything about technicians. Describe a different operation — its resources, skills and rules — and the same system should schedule that too.