Hi everyone
I am a student building software in my free time, for the last few weeks I have been building agonistes.app. What started with an llm wrapper for coaching, just like all the others, is now a JEV workflow with haiku 5.5 for text to the user but JEV chooses workouts based on science and an entire database of trainings. Currently a few friends are testing and it works well, but not yet good enough for long term planning with real competitive athletes. I just had a few questions:
- Has anyone used JEV or another system one model for AI coaching? If so did you finetune/ reinforcement train it (probably only possible with open source version)
-Currently it just works with intervals.icu API, would need connector later, anyone who knows how to get it with garmin, wahoo, whoop,.. Which ones are possible?
-Would coaches be interested in a coaching version, they would need the API of their athletes, use the platform which is like trainingpeaks reverse engineered, but with the ai build in for data analyzing, a copilot and a real building engine?
Thanks a lot in advance!
This is actually quite interesting.
I had a similar idea a few weeks ago to use JEV to quickly and cheaply comb through the massive amount of personal training data I’ve accumulated over the years. I ran a few experiments to see what it could do, with the biggest one being whether it could reliably predict past race performances based on training data leading up to those races.
Unfortunately, the results weren’t particularly promising. My suspicion is that the original training data used by TypeSafe AI didn’t include much exercise or endurance training data, because JEV’s results were pretty lacklustre and significantly behind the frontier LLMs I tested it against.
I’d be really interested to hear how you’ve found JEV performs when analyzing training data, and whether you’ve had better results.
I’d also love to see a sports science research group get their hands on training histories from thousands or millions of athletes and use that dataset to train or fine-tune something like JEV. I suspect that’s where things could get really interesting, potentially even helping resolve some of the big open questions in endurance exercise research.
IntervalCoach founder here, so I’m biased, but we’ve had Jev in production for a few weeks and the experience lines up with what @RunTK.com describes.
The short version: we use it as a classifier, not as a coach. It never analyses training data, picks workouts or writes anything an athlete reads. Where it earns its place is narrow: a decision with a fixed set of answers (a choice, a yes/no, a score), made on free text that varies a lot by language or by who wrote it, where a wrong answer actually costs the athlete something. Reading what someone means by a chat message, or which of two similar planned sessions an activity actually completed, are the kind of thing it’s good at.
Hi thanks for the reply!
Currently its good at qualifying, choosing effort for LLM. I set up a sandbox repo and let agents do research and experiments for days, some promising results, also they got too fixated on safety and medical. Still need to skim through all of it.
Not used it for predicting, could be useful, maybe even together with a real own ML model.
The coaching engine is now hybrid LLM+Jev, will try to use obsidian to store training libraries and coaching knowledge better, maybe jev can choose the obsidian paths (obsidian is just an md file structure, no second brain :))
Also thought about this: using Laya, the open source local version and finetuning it for endurance coaching, or trying to build an own system one model, maybe if my university let’s me do it for a project, to not waste too much time on it.
Happy to hear your thoughts and experiences about it.
Will post blogs with findings on agonistes.app.
Also watching beprimed, they are also experimenting, maybe try to reverse engineer it a bit.
Thx for your reply!
Classifier is also in our engine, doing more research right now, see reply to @RunTK.com.