Serving

The infrastructure that runs models in production: batching, caching, routing, autoscaling. Where 'it works in the demo' meets physics and unit economics.