I’ve tried all the AI fitness coaching out there. My watch can give me hardcore race-prep advice, and my treadmill can give me plenty of cheerleading, but none of them really know me. They don’t remind me to set out my shoes before a long run, or that I always forget the anti-friction cream. And I can’t tell them when I’m tired, busy, stressed, or injured and expect the plan to actually change.
I wanted a coach that adapts to my real life.
So I built one.
This talk walks through my personal AI running coach, an end-to-end system I use almost every day:
- The data problem: merging workouts from Strava, Coros, and Peloton, plus training notes from Slack; building small DIY “unofficial APIs” for platforms that didn’t expose what I needed
- The architecture: Slack bot -> AWS (EventBridge, DynamoDB, Lambda) → unified training dataset → Groq LLM
- The AI model: using the open gpt-oss-120b model running on Groq for fast, inexpensive inference
- The prompt engineering: getting the model to behave like a real running coach instead of a generic motivator
- The infrastructure: everything provisioned and automated, inexpensive enough to run continuously
- The lessons: what surprised me, what misfired, and what actually changed my training
Although this project is about running, the bigger idea is more important: we’re entering a world where individuals can build their own personal software, powered by their own data, shaped around their own routines, and not dictated by a generic app.
It’s a glimpse of something new: handcrafted, user-owned, deeply personal AI.
If you’ve ever wanted an assistant that truly understands your life, this talk shows how close that reality already is.
