Personal context.
Connected intelligence.

The starting point

Understanding training meant moving between activities, strength records, recovery information and personal notes. The founder built a personal AI Coach to bring those sources together and explore questions about progress, completed rides and the next training choice.

What was built

The project contains source adapters, stored reports, Python analytics, a desktop dashboard, an authenticated API, a read-only iPhone companion and a bounded read-only MCP adapter. Local Ollama models explain supplied evidence.

• Progress views across cycling, strength, recovery and body-composition evidence.
• Stored ride reports and comparison facts.
• Goal and availability inputs, with provisional training options.
• Approved notes retrieved to add relevant context.

The architecture decision

Structured records are held in hosted Supabase. Python computes metrics and rule-based options. Local Ollama inference adds natural-language explanations; Qdrant supports retrieval of approved notes.

This is a hybrid design. Local model inference does not mean that every application record stays on the Mac.

Controls that matter

The source separates facts from explanation, records evidence fingerprints and prompt versions, checks for some unsupported numerical claims and withholds failed explanations. Missing information is represented as missing, rather than silently treated as zero.

These are engineering controls intended to reduce unsupported output. They do not establish that every explanation is correct or that hallucinations have been eliminated.

What remains to learn

The implementation is a personal prototype, not a finished commercial service. Physical-device acceptance, longer-running operation and broader evaluation of model usefulness remain important areas of work.

The transferable lesson is architectural: give each component a clear role, retain evidence and make the human review point explicit.