For most of the time I've been building Gym Companion, there was no AI in it at all — and I was in no hurry to change that. Sometime last year every app I opened had sprouted a chat bubble in the corner, ready to "coach" me. Most of what came out of them was the same confident, generic advice you'd get from any fitness influencer with a ring light, except now it could also quietly make things up. I didn't want to ship that.
Here's what actually bothered me. Advice about training and nutrition isn't short on opinions — it's short on the ones that survive contact with the evidence. How many reps build muscle, how much weekly volume is actually enough, how many calories you really need: there are decades of research on all of it, sitting right next to whatever happens to be trending this month. A model trained on the whole internet will hand you both in the same self-assured tone and leave you to work out which is which. Bolting that onto a workout log felt less like a feature and more like a liability.
So I turned the problem around. Anything that can be calculated, I calculate — no AI involved. Your calorie and protein targets come from a formula that's been in the sports-science literature for decades, run against your own body and goal. How hard you've actually trained each muscle group over the last two weeks is counted, set by set. Your strength trend is plain arithmetic on the weights you lifted. None of that is a matter of opinion, so I don't ask a language model for one. The app just does the math, on your device.
That leaves a narrower, more honest job for the AI: look at your numbers — not the internet's — and reason about them the way a good coach would. I hand it a tight brief built on the boring, well-established principles: the rep ranges that genuinely drive muscle growth, the weekly volume where most people actually progress, the fact that a muscle trained twice a week beats once. And I tell it what not to do — no pushing someone training for general health to chase a heavier bar every single week. It reasons inside those guardrails, on real data.
It suggests. It never decides.
That distinction matters more than anything else here. Every adjustment the app arrives at — add a set here, nudge the weight there — is a suggestion you tap to accept or ignore. And I'll be straight about why it works that way: this is still a language model, not an oracle that's been fact-checked line by line against a research library. It's grounded in sound principles and it's reasoning about your data, but it can still be wrong — and I'd rather you hold the final say than pretend I've built something I haven't. Keeping a person in the loop isn't a limitation I'm apologising for. It's the design.
I didn't want AI in the app because it would look modern on a feature list. I wanted it to earn its place the same way voice did — by doing one thing genuinely well and staying quiet about the rest. Calculate what can be measured. Reason carefully about the rest. And hand the final call back to the person actually lifting the weights.
More in two to three weeks.