Speeding up the art of ML model deployment
The value of redeployed models
Modelling and eng part to ship it ⇒ what happens when we have new features, or drift, etc
Any model that you train today will be obsolete tomorrow. By how much will it be obsolete?
Domain shift
Bottleneck when it comes to production
Skill set for regular ML deployments
Need to be 10x DS?
very operational work to bring value
Improving model release processes
automate the majority of the pipeline
another trick to de-risk the pipeline is to leverage shadow mode
deploy our prod model and also shadow, to observe how it behaves in production
Schedule it
