AI coaching that speaks human, not numbers
For aspiring triathletes who want to step up their game, meaningful feedback beats arbitrary metrics. The service explores how AI can provide coaching that compares you to yourself—'How are you pacing this segment compared to last time?'—rather than just generic performance data.
Built for hobbyists serious about improvement, the system delivers a complete coaching experience: preparation beforehand, real-time guidance during training, and analysis afterward. This includes coaching across running, cycling, swimming, strength, and conditioning, with a strong emphasis on recovery.
The technical challenge went beyond standard AI implementation, requiring solutions for: custom underwater stroke detection to bypass SDK limitations; offline audio coaching for swim-specific headphones; and a non-negotiable, privacy-first architecture where users own their health data.
Solving for SDK Gaps and Offline Training
Building in Swift for iOS and watchOS immediately revealed SDK constraints. For instance, underwater stroke detection required custom algorithms to predict swimming phases and distinguish between race starts and turn transitions—functionality not available natively.
The 'coach in the moment' experience relied on offline audio synthesis. To support devices like the Shokz OpenSwim Pro headphones, the system pre-generates personalized coaching audio, ensuring contextual guidance is available underwater or without connectivity.
Technical stack and integration:
• Swift for native iOS and WatchOS development
• Claude and Gemini APIs for AI analysis and coaching generation
• iCloud for cross-platform data storage and sync
• Custom algorithms for underwater activity detection
• Offline audio synthesis for uninterrupted coaching delivery
• Privacy-First Architecture: User-controlled data, not for mining or sale
The challenge wasn't just making the technology work; it was making every technical decision serve a specific training need while protecting sensitive health data.
Human-friendly analysis over raw metrics
Instead of overwhelming athletes with raw numbers, the AI translates performance into meaningful, conversational insights. A 'ghost mode' allows athletes to train against their own previous performances in real-time. This addresses the core user need: measuring progress against themselves, not arbitrary benchmarks.
An insight like, 'You maintained consistent pacing better than your last three sessions,' carries more training value than a simple list of split times.
Closed beta testing insights:
• Comparative language is far more motivating than absolute metrics.
• Real-time audio feedback is crucial when visual attention is impossible (e.g., swimming).
• Post-workout analysis is most valuable when it identifies multi-session patterns, not just one-off results.
• Recovery guidance (or lack thereof) was a critical factor in training adherence and injury prevention.
The system learns communication preferences, adapting its style while keeping number-heavy data available on-demand for those who want to dive deeper.
Privacy-first architecture: Building trust
Sensitive health and performance data requires an ethical, privacy-first architecture. This was a non-negotiable principle: workout data belongs to the user, not the app, and is never mined or sold.
This principle drove technical decisions at every level:
• iCloud Storage: Data is stored in the user's personal iCloud, under their control.
• AI & Data Boundaries: AI processing is designed to respect data boundaries and limit exposure.
• Full Transparency: The app is clear about how and why coaching insights are generated.
Trust is the foundation of any coaching relationship, whether human or AI. This system is designed as a supportive, private partner that respects both the user's ambitions and their data.