03 / AI

AI, in practice

Building with models, reviewing their work and understanding where they need help.

My interest in AI is tied to the work I can do with it. I use it while building software, and I’m building tools of my own to make that process easier to follow. That gives me practical questions to work through: which model to ask, how to check the answer, and how to keep a project coherent as the work moves between conversations.

My background in delivery influences how I approach this. I want to understand what has been decided, what has actually been checked, and who or what needs to act next. Those questions remain useful when some of the work is being done by a coding agent.

01 / Model review

Keep the question and the second opinion together.

In Dispatch Console, I choose a model for the initial question and can ask another to review the response. Keeping the exchanges together makes it easier to see where the answers differ and decide what needs further investigation. The cost of each call is visible alongside the work.

I’m interested in reviews that identify a specific problem or assumption I can check. Agreement between models can be reassuring, but I still need to examine the result against the task.

Inside Dispatch Console

02 / Coding agents

Make the work and the handover visible.

AI Workbench is my exploration of a local workspace for coding agents, with writer and reviewer roles around an item of work. The project board and individual sessions give me a way to follow progress without piecing the whole story together from separate terminals.

I’m developing it around the questions that arise between sessions: what context should carry forward, what a reviewer has asked for, and whether those points have been addressed before the next piece of work begins.

Inside AI Workbench

03 / Useful output

Bring the answer back to the situation.

MARIAGE gives me a different kind of problem. A pairing suggestion needs to make sense with the food and wine available to someone at a restaurant, and the explanation needs to help them choose. The wording and the way the options are presented are part of whether the application is useful.

Across these projects, I keep returning to the same practical check: can I explain why this result is useful, and what would I look at to find out whether it is wrong?

Inside MARIAGE

AI Workbench development view with sample data — Development view · sample dataOpen full size (opens in a new tab)

I’m happy to compare notes on using AI in software delivery, model review or the tools around coding agents.

Get in touch