Hi all,
I'm building a cycling coaching app — direct integration, everything on our own servers. All the training analysis is produced by a deterministic engine (TSS, CTL/ATL/TSB, power/HR zones); no AI model produces the analysis, and we never train, fine-tune, or build embeddings on Strava data.
My question is specifically about Section 5.3 (AI).
To phrase the coaching reply in natural language for that same athlete, one option is to pass only the derived metrics our engine already computed (e.g., "TSS 65, IF 0.82, zone 3") — never raw Strava data — to an LLM (OpenAI, under no-training / DPA terms), strictly at inference time, only to turn our numbers into a human-sounding message.
Section 5.3 is worded broadly ("operation of any AI Application... any data derived from... ingestion into a context window").
1) Does this prohibit inference-only use of derived metrics, or is it aimed at training / RAG / embeddings?
2) If it does prohibit it, would an approach where the LLM only ever receives opaque placeholders (e.g., {{TSS}}, {{zone}}) and our engine substitutes the real values afterward — so no Strava-derived data reaches the model — be compliant?
Trying to get this exactly right before building further. Thanks for any guidance from the team or other devs.