Builds

T1D Clinical Audit Tool for Claude Desktop

Watching technical tutorials on YouTube usually ends the same frustrating way. You pause the video, copy code off the screen by eye, alt-tab to your editor, realise your setup does not match the presenter's, and lose the thread entirely. Most video collateral is sitting there inert. Someone watches it, and that is the end of the transaction. Prompt Track changes that. It turns ordinary YouTube videos into structured, interactive courseware by attaching a timestamped sidecar of authored cues directly to the footage.

The first iteration of Omni-Endo AI showed me that handing Glooko data to an LLM could produce some interesting ideas and give some great insights. But it required a manual copy-and-paste handoff. This was done to avoid LLM tie-in and also to avoid API costs.

Precision asset extraction from flat-colour images — no AI, no server, no signup.

The Knight's Tour is a chess problem that has been keeping mathematicians and puzzle enthusiasts occupied for roughly a thousand years. The challenge is to move a knight across a board so that it visits every square exactly once. It is one of those problems that sounds straightforward until you actually try it. This is a colourful variation on that idea, with an added constraint: instead of visiting every square, you must visit every colour-coded region exactly once, in valid knight moves, with no backtracking.

This started with a straightforward frustration. The historical data sitting in Glooko, particularly from the Omnipod 5, is far richer than anything the manufacturer's own app surfaces in a useful way. The question eventually became: what happens if you hand that data to a large language model and ask it to behave like an endocrinologist?

This started with a simple question: when exactly did I write that? LinkedIn's own timestamps are vague at best and drift into "a while ago" fairly quickly, which is not particularly useful if you want to know the precise date of something you posted a while back. So I went digging, and it turns out the answer was sitting in the URL the whole time.

This one started with LinkedIn. They have a daily logic puzzle called Queens and it is genuinely good fun. One a day, however, is not quite enough for a certain type of person, and the leaderboard had developed a noticeable population of individuals claiming to solve nine by nine boards in four seconds. Whether that is humanly possible is left as an exercise for the reader. So I built my own version. Queens With Friends is an open-source logic puzzle engine based on the same core concept: place one queen in every row, column, and colour region, with no two queens touching, including diagonally. The interesting part is what sits around that core. You can build your own boards, generate random ones at various difficulty levels, and share them directly with friends via a link that contains the entire board state. No server required.

This one started with a job application. I was looking into Temporal as part of the process and, as tends to happen, the learning took on a life of its own. The result was a playable game rather than a tidy tutorial project, which probably says something about how my brain works.

This is a single-page browser tool built around a concept borrowed from classic espionage: the newspaper code. The idea is that a secret message is hidden inside a perfectly ordinary piece of public text, and only someone with the right key can extract it. No invisible ink required.

Fitness apps figured out years ago that gamification gets people engaged with their data in a way that raw numbers never will. Diabetes apps, largely, have not caught up. Glucose Grand Prix is a small attempt to show what that could look like.

This one started in a garden, with a seven-year-old and a perfectly reasonable question. My niece asked whether it was true that if she dug straight down she would end up in Australia. I told her it would be close, but probably somewhere between Australia and New Zealand, which is the sort of answer that satisfies most people. She was not most people. She pointed at the ground a few feet away and asked: "Where would that go to?"