1Context
A live-event platform hit its ceiling
Concerts, sports and interactive shows for audiences that spike within minutes. Half of the viewers watch on mobile, often on unstable networks.
2Challenge
Latency, drops and license bills
- Glass-to-glass latency grew at peak load; chat and reactions fell out of sync with the picture.
- Viewers on weak mobile networks dropped instead of degrading gracefully.
- Per-stream licensing for the off-the-shelf codec stack scaled with the business.
- Moderation relied on people watching every stream.
3What we build
An in-house video pipeline
- Encoder tuning and codec-level optimization for mobile: smaller keyframe intervals where it matters, smarter rate control, lower bitrate at the same perceived quality.
- Adaptive bitrate ladder built from real device and network data, not defaults.
- Edge distribution with regional ingest and failover.
- ML moderation and highlight detection running on the same stream, with human review for edge cases.
- # stack · models · guardrails
- interactive_rooms: WebRTC over an SFU (mediasoup / LiveKit)
- broadcast_scale: LL-HLS with regional ingest and CDN egress
- encoder: tuned rate control and GOP structure for mobile networks; codec choice per licensing constraints
- abr: per-title bitrate ladder derived from real device and network telemetry
- moderation: CV models on the live stream, human review queue for edge cases
- resilience: multi-region ingest with automatic failover
4Outcome
One pipeline for stadiums and small rooms
Stable low latency at peak, fewer mobile drops, licensing cost decoupled from audience size — and the same system now runs interactive rooms with a handful of participants.
<2s
glass-to-glass at peak
−35%
bitrate at the same quality
mobile-tuned codec
$0
license cost per stream
decoupled from audience
100%
of streams moderated in real time
ML + human review
−40%
cost per viewer-hour
license + CDN egress, at the same audience