Athlete Space.
AI-native endurance-training platform. Multi-agent LLM coaching, real-time pose-estimation form analysis (ONNX MoveNet), and integrated wearables, grounded in objective training-load science (CTL / ATL / TSB). Not another chat wrapper.
The problem
Endurance coaching is rich signal, poor tooling.
Endurance athletes already produce some of the cleanest, densest objective signal of any consumer domain: HR, power, pace, RPE, sleep, HRV, training-load history. What they have to work with on top of that data is usually a static plan and a human coach they see once a week.
The approach
Specialist agents on an objective substrate.
Athlete Space pairs a deterministic training-load pipeline with a multi-agent LLM coach. The LLM layer doesn't invent physiology. It reasons about it. Every interaction passes through typed contracts, verification, and a reply-synthesis layer.
Architecture highlights
Pipeline
Deterministic plan engine.
A seven-stage planner (macro plan → philosophy → week structure → volume allocation → session templates → session text → calendar) with a single LLM call per stage and deterministic fallbacks.
Orchestration
Multi-agent LLM coaching.
A production multi-agent system designed with athletes and coaches as the end users. Planner, Advisor, Evaluator, Verification, Reply Synthesizer, and more — each a typed agent with structured outputs, retrieval, and evals.
Governance
Eval loop & rollback.
Feature-flagged rollout, regression suite, behavioural eval harness, and an auditable rollback plan. V5 shipped on a measured, reversible path.
Stack
Backend.
AI surface.
Surfaces
Why I'm building this
I'm the user.
Sub-2:30 marathon. Ironman champion. A decade of structured endurance training while leading AI at Google, Amazon, and now Wizard AI. Athlete Space is the coach I wished I had, built on a substrate I trust.