ML in production

Training a model is the first third of the work. The rest is: expose it, freeze the environment, put it somewhere other people can hit it.

ML-in-production is a complete path:

1. Model — Keras / TensorFlow, including pretrained backbones
2. API — Flask, so the model is a HTTP interface rather than a notebook
3. Container — Docker, so the next machine does not become a dependency puzzle
4. Deploy — Heroku as the public endpoint

The accompanying article is Machine Learning in production: Keras, Flask, Docker and Heroku. A follow-up covers running multiple Flask apps behind Nginx and Gunicorn.

This is still the shape I use when someone asks “how do we ship this?” — swap the cloud, keep the seams.