Really interesting thread. I work with AI as a developer and took a similar approach to this problem as coding agents do.
The core idea: don’t feed data into the model. Give the AI agent tools to search and retrieve only what’s relevant to the current question. No massive dumps, no context window degradation.
Add a visualization layer on top — the user sees the same charts the AI is reasoning from. If it says “your decoupling was 8.7%,” you can see the power vs HR drift yourself. Kills the black-box problem and makes hallucinations immediately obvious.
Built this into an app (Thread) that integrates with intervals.icu.