Coffee Compass

Challenge

The Fellow Aiden brews beautifully once a profile is dialed in, but dialing in is a craft, and the easy path is to accept a roaster's or an app's preset that was never built for your palate. I spent a year and a half dialing profiles with a custom GPT, and by the end I had what I needed to run the whole thing myself: a pile of profiles, a clear sense of my taste, and the infrastructure to host it. I wanted a tool that learns what I actually like, helps me get there quickly, and remembers every cup so the next one starts smarter.

Build

Coffee Compass keeps a library of every coffee I've tried and generates Aiden profiles grounded in three things: real brewing principles, a model of my taste, and my own dialed-in recipes, so every profile reasons from my own hand and history. It talks to the Aiden in both directions. I can push a finished profile to the brewer and get a brew.link QR for it, and I can pull every profile back down into a durable archive that outlives the device's 14-profile limit.

A "My Aiden" view shows the whole collection by state, what's loaded on the brewer now, what's queued to move, and what's kept off-device, with one-click promotion. A Brew Log keeps the running story of each coffee: the tasting notes I write and the dial-in changes I make, and those notes feed the next regeneration, so it tightens over time.

Outcome

A brew companion that gets more like me the more I use it. It lands conventional coffees close on the first pass and gives a strong starting point on the rest, then sharpens through feedback.

I went back and measured this against my own dialed-in profiles. On the coffees I dialed from scratch, the model predicts my ratio, my pour structure, and my bloom temperature closely even when it has never seen the recipe. On Fellow Drops, where the roaster publishes a recipe to compare against, the version grounded in my own profiles lands closer to my dialed-in settings than the roaster's preset does. The from-scratch sample is small and I report it as such, and bloom still carries judgment that no metadata predicts, so the model gives a strong starting point and the tasting loop closes the last few degrees. The longer I use it, the more it reflects my own preferences, and that is the whole point.

I am sharing it as a build-it-yourself PRD. Anyone with a Fellow Aiden can hand the brief to their own AI agent and build their own at-home profile maker: one that learns their own preferences, or takes a roaster's preset plus their own taste as a starting point to dial in from.

Software is increasingly disposable, so the blueprint is the thing worth sharing, the part that lets each person build their own.