Presentation: From POC to Production in Minimal Time - Avoiding Pain in ML Projects

“So how soon can this go live?” It can be a chilling question because you know that whatever answer you give, there’ll be a business need to get delivered sooner and with fewer resources than you need.  Turning an AI proof of concept into a production ready, deployable system can be a world of pain, especially if different parts of the puzzle are fulfilled by different teams. When promised data doesn’t appear and timelines and scope creeps what can you do?

I’ll talk you through one such project: from the initial pitch and how that changed to the agreed project deliverables, the first AI model and a very clunky web demo, dealing with the extensive missing data and creating an automation pipeline to deal with it, getting a tensorflow based image classifier working in docker with a slick front end and continuously updating and deploying itself using codeship and AWS fargate. For each step, I’ll go into the technical detail so that whichever part of this puzzle you’re missing, you will be able to fill in the gaps and put something similar together yourself.

Key takeaways:

• Setting up a data pipeline so that you can feed your models
• Creating an api accessible ML model
• Docker with GPUs and if you need it
• Adding a demo suitable for clients not data scientists
• Making production quality ML

How to put everything together and set up a continuous delivery pipeline for MLs models using docker, or staged deliveries using AWS fargate.