philippe-ths/ai-running-coach

Running Coach — an AI coach for your Strava runs (single-user MVP)

Running Coach connects to Strava, deeply processes each run into deterministic training signals, and turns those signals into an ongoing relationship with an LLM coach: short, human, opinionated coaching that arrives after you run and remembers you over time. It runs locally via docker compose, or as a deployed single-user instance on Railway (backend) and Vercel (frontend) — see Production deployment.

Most Strava-based tools surface raw stats but little actionable insight — they tell you what happened, not what to do next. This project bridges that gap: a correct, auditable deterministic substrate (metrics, blocks, training load) under an LLM coaching layer that has a voice, a point of view, and durable memory — while never letting personality override the measured data or a safety floor.

What it does

• Deep run analysis. Pulls activity streams from Strava and computes effort, pace variability, HR drift, time-in-zones (binned against your own Strava HR zones), interval structure (from your recorded laps), stops, efficiency, risk flags, and a confidence read.
• An AI coach, not a report. After a run you get an immediate, input-free message, then a fuller coaching turn once you have had a moment (or once you reply) — delivered over Telegram with one-tap RPE/pain buttons so it never blocks on input. The coach writes human prose first and emits structure only as a thin tail.
• Personalization you control. Declare the coach's Voice (warmth / humor / directness / energy, plus presets), its Coaching stance (a training school of thought + two emphasis dials), and upload your own coaching materials. These flex how the coach talks and what it foregrounds — never the facts, your goal, or the safety floor.
• Memory that compounds. A durable, split memory (auditable deterministic facts + an LLM-written relationship narrative) plus a learning loop t