Engineering notes from building AudioRun
2026-08-30
Mid-run, the announcements aren’t a feature — they’re the whole interface. How spoken stats are assembled clause by clause, how state changes, stats and the AI coach share one voice, and how the app ducks Spotify, lets Audible pause, and stays out of a Garmin watch’s way.
Read the post →2026-08-23
The voice coach used to speak on a fixed schedule. Now it also reacts to heart rate zones, climbs and fading pace. A 2 km test run where I started too hard up a hill shows why a distance-based coach would have arrived too late.
Read the post →2026-07-21
Beta testers wanted Strava-style comparisons on familiar ground. AudioRun clusters your usual loops on-device from GPS overlap: fingerprints, a bearing gate for out-and-backs, and shared stretches when two routes only partly overlap.
Read the post →2026-07-17
A coach that speaks mid-run has to be fast, short, honest about numbers, and silent about things it can't know — and you can't test that exhaustively on real legs. How AudioRun's coaching is evaluated with simulated runs replayed through the production pipeline, and judged reply-by-reply by Claude Opus 4.8.
Read the post →