"AI coaching" has become a sticker. Almost every training app now claims it. Peel most of them open and you find the same thing: a single general-purpose chatbot with a cycling-flavoured prompt bolted on top. It talks a good game. It doesn't actually model anything.
CoachedBy is built the other way around. Underneath the conversation are three genuinely different kinds of AI, each doing the job it's actually good at — and they're wired together so they reinforce each other instead of bluffing.
Three engines, not one chatbot
1. Physics- and physiology-based models
This is the part a language model can't fake. Your critical power (CP), your W' (the matchbook you can spend above CP), your Pmax, your CTL/ATL/TSB fitness–fatigue balance, your full power–duration curve — these come from validated sports-science models, computed from your real efforts. Not estimated from "intermediate cyclist, four hours a week." When CoachedBy says your threshold is 285 W, that's a number a model fit to your data — open your physiological twin and you can see it, and tune it by hand.
These models also simulate. Ask for a pacing plan up a 40-minute climb and the platform runs your power against a W'-balance model — an actual physical computation of how deep you'd go and whether you'd blow up — not a guess that happens to sound plausible.
2. Machine learning that calibrates to you
Raw models have to be tuned to a real human. That's the data-driven layer: CoachedBy fits your CP and W' as new efforts land, cleans noisy power, models how your heart responds to load, and positions your levels against a large cohort of real riders instead of a textbook table. The numbers move as you train. They're learned from your data — and thousands of others' — not typed in once and left to rot.
3. Generative AI for the conversation
This is the layer most apps stop at, and it's the one you talk to. CoachedBy uses a top generative model to explain what the other two engines found, build your week, answer your questions in plain language, and adapt when life gets in the way. The crucial difference: it reasons over the outputs of the real models, on your real data. The generative AI is the interface — not the thing inventing your physiology.
Why "all in one place" is the whole point
Pick almost any other tool and you get exactly one of these:
- A power-analytics platform with great models — but no coach, no conversation, no plan that adapts to you.
- A chatbot that talks fluently about training — but with no validated model of your body underneath.
- A coaching app with a human — but generic, one-size plans.
CoachedBy is the first platform to put all three together and connect them: real physiological models, calibrated to you by machine learning, with a generative coach reasoning on top. Each engine covers the others' blind spots — the generative AI can't hallucinate your CP, because a model already computed it; the model can't talk to you, so the generative layer does.
Built by people who've done this at the top
This isn't textbook code. The modeling under CoachedBy is WorldTour-level — the same approaches used inside professional cycling programs to model durability, fatigue, and pacing — built by WorldTour data scientists who've done exactly this work for pro teams. What used to live in a performance department's private toolkit now runs on your own rides.
What it means for you
- Your targets are actually yours — computed from your power, not a category you got dropped into.
- The analysis is physically grounded — when the platform says you'll fade at 38 minutes, a model said it, not a vibe.
- Pacing and projections are real computations — simulated, not improvised.
- The advice stays consistent — because a methodology drives it, not whatever the chat felt like today.
A chatbot can describe coaching. CoachedBy computes it. That's the difference between something that sounds like a coach and something that is one.