Software Engineer
Moonshot (67 Labs, YC S26) · REMOTE (CONTRACT)
Voice, audio and speaker-ID on iOS for a Y Combinator S26 startup.
- 44 → 0
- FALSE ACTIVATIONS
- 9.1×
- FASTER SPEAKER-ID
- 12×
- LESS BANDWIDTH
- Built a media-aware voice activation gate using iOS playback signals, cutting false activations from background Reels 44 → 0 across 1,000 simulated rooms; 0 unwanted responses on 729 production turns, with no on-device model.
- Optimized WeSpeaker speaker-ID with FP16: 74.4 ms → 8.2 ms CPU time (9.1× faster), identical decisions across 6,514 test cases.
- Switched server-to-iPhone audio from PCM to Opus: 384 → 32 kbps (12× less bandwidth) on constrained networks.
- Eliminated a 4–5 min re-onboarding for returning users after reinstall by reconciling local and server state.
- Shipped iOS ShazamKit music ID end to end: server-triggered on-device capture, ≤12 s listen window, Apple Music result card, zero audio retention.