A live pipeline has five stages: ingest from the event feed and the video, feature extraction (vision-model outputs and structured events), detection of a moment worth acting on, generation of the brief or overlay, and delivery into the production system. A sub-second budget end to end leaves roughly two hundred milliseconds for generation once the rest is accounted for.

Hosted frontier models rarely get under that even before the network round-trip, and their tail latency under load is unpredictable. That is why the generation stage is the one that moves to the edge.

Small language models, quantised and served on GPUs next to the ingest, handle the time-critical generation: a one-line overlay, a short narrative brief in a fixed format. They are prompted with structured context the pipeline has already assembled, so they are completing a template as much as writing. Kept narrow, they are reliable.

The frontier models still have a job: long-form summaries, multilingual output, and anything editorial that can wait thirty seconds. Routing by deadline rather than by task type is the design principle.

Fusing a video feed, an event stream, and a statistics API into a single timeline with consistent timestamps is most of the engineering. Every source has its own clock, its own lag, and its own failure modes. A low-latency event bus with explicit event-time semantics, and a fusion service that tolerates late data without duplicating moments, is the foundation everything else stands on.

We size for concurrency from the start. Matchday peaks with a dozen concurrent fixtures are where the pipeline earns its keep; a single-match demo proves nothing.

The metric that counts is how many generated briefs a senior editor accepts without changes. We build the evaluation harness around a panel of editors scoring real output, and no model or prompt change reaches production without the score holding.

Acceptance rates in the high eighties to low nineties percent are achievable for narrow formats. They are not achievable for open-ended writing at these latencies, which is why the formats are narrow.

Matchday operations need the same discipline as any production system: dashboards on end-to-end latency per stage, alerting on tail latency rather than averages, and a kill switch that drops back to manual production instantly. The production team has to trust that the system fails quiet, not loud.

What we tell clients: fix the budget first, then design backwards from delivery. Anything that cannot make the budget goes on the slow path, and the slow path is still valuable.