Hey everyone,
I’m Luiz Tessarolli, a senior software engineer from Brazil, and amateur MTB racer who stopped training seriously for a couple of years to focus on career and family.
Now that I wanted to get back on training, and there is a plethora of AI Coach’s and Apps and they all claim to be the “best tool whatsoever”, but then again, its an LLM wrapped with some custom instructions, that may or may not fix LLMs biggest problems, hallucinations, memory, and incoherence between sessions. I’m not saying they are bad in any way, I’m saying is that I don’t know how they are made and what they actually want.
So I built one for myself. With the rules and governances that would make me trust an LLM generated output a little better, instead of the “trust-me bro” vibe, Prova Endurance justifies every claim from a previously indexed and approved scientific paper corpus (currently with 110 papers on it), with the quotes and the papers to prove where a suggested claim came from, or, if its from its own judgment, transparently said so, and all suggestions are treated as only that, suggestions, until you review and approve it yourself.
The MCP server did a good job, but I wanted to see some charts, then I added a nice UI to it.
It turned out too good to keep completely private, so I’m opening it to everyone:
The following is an LLM generated description of Prova Endurance:
The basic idea is a little different from “an AI coach.”
Your Claude, ChatGPT, Gemini or Le Chat is still the thing you talk to. Prova Endurance connects to it through MCP. The AI handles the conversation and reasoning; the Prova server stores your training history and profile, calculates things like training load, fitness and zones, and checks proposed plans against a set of training rules.
That distinction matters to me.
LLMs can be very convincing while still getting numbers wrong or losing context over a long conversation. I didn’t want the model inventing training data or quietly changing things.
The server currently has a research corpus of 110 notes based on published papers. When the coach gives a reason for a session or plan, it either cites that research or explicitly says that the recommendation is the coach’s judgement.
There is also a “judge” on the server. Plans are checked for things like load progression, weekly load, spacing between hard sessions, recovery weeks, easy-session distribution and taper structure.
And when the AI proposes a change, it doesn’t get to silently apply it. The athlete accepts or declines it, and the decision is recorded.
There is one important exception:
Once a plan was approved and is running, the next week is filled in automatically as the weekly prescription and emailed to the athlete. You can then change it through a proposal.
There is also a Coach tab in the web app where you can inspect proposals, decisions, observations, profile changes, emails and the tool calls the AI made to your data. You can also revoke a connected AI app.
For data, you can connect intervals.icu, import a Garmin Connect export, or upload FIT/GPX/ZIP files. Planned workouts can be downloaded as FIT files, and with intervals.icu connected they can be published to your calendar. (Mine syncs to Garmin devices automatically)
I should also be upfront about where this is today.
This is early. I’m currently the only athlete using it. I’ve mostly used it for my own Cycling/Strength trainings.
I’m not presenting this as a replacement for coaches, and I’m definitely not claiming it has already proven anything at scale.
I built it because I enjoy building software, I already have the infrastructure and AI subscriptions, and I wanted to solve a problem I personally had.
If you’re curious, give it a try and tell me about it.
