Anonymised under NDA
Real-time match intelligence on live sports data
Sports & Media · A sports and media operator
Real-time event detection and narrative briefs across a live match feed, fusing vision models, structured event data, and language models.
The problem
- Manual production teams could not keep pace with the volume of moments worth flagging across concurrent fixtures.
- Existing tooling treated video, event-stream data, and statistics as separate worlds with no unified narrative.
- Latency budgets for broadcast overlays were sub-second, far below what hosted LLM round-trips could meet.
The architecture
- Streaming ingest of event-stream feeds and frame-level vision-model outputs into a low-latency event bus.
- Edge-deployed compact language models for narrative generation, with hosted frontier models reserved for higher-stakes briefs.
- Evaluation harness scoring briefs against a panel of senior editors before any model change reaches production.
KafkaVision TransformersLlamaOpenAIRedis StreamsKubernetes
The outcome
- Production teams could focus on storytelling instead of monitoring; the system surfaced moments they would otherwise have missed.
- Multiple language outputs ran in parallel without expanding the editorial team.
Services drawn on
A similar problem on your roadmap?
A Discovery Sprint is the fastest path to a scoped, costed proposal.