Two decades inside live broadcast. Now I build the AI that runs in it.

Computer vision for broadcast

Models tuned for footage nobody cleaned up first: compression artefacts, moving cameras, floodlights, bodies in the way.

• Mask-level segmentation
• Occlusion and camera motion
• Measured accuracy and throughput

Real-time video pipelines

NDI, SRT and file pipelines that hold their frame rate for a full match, not for a thirty-second demo clip.

• GPU decode and resize strategy
• Latency budgets you can hold to
• Logs and metrics from day one

Timing and data systems

The systems that put numbers on screen. I’ve run them at championship rounds where a wrong result is visible to everyone watching.

• Interfaces operators can drive under pressure
• SQL-backed results that survive a crash
• Clean feeds for graphics and web

Broadcast

Software

AI / Vision

Cloud / DevOps

I speak both languages

Vision engineers usually haven’t sat in an OB truck. Broadcast engineers usually can’t write a segmentation pipeline. I’ve done twenty years of one and three of the other, and the translation between them is where most projects stall.

I’ve been there when it broke

Championship rounds, foreign venues, kit that arrived late, a feed that dropped twenty minutes before transmission. Two decades of show days taught me to design for the failure first and the happy path second.

I finish the integration

Most AI work dies between “the model works” and “it’s on air”. I don’t hand over a notebook — I deliver the pipeline: ingest, process, validate, publish, monitor.

I build for the person operating it

I’ve written the software and then stood behind the operator using it live. If it isn’t obvious to someone with a director shouting in their ear, it isn’t finished.