Do you have any models in production, or do you plan to have any?

General Discussion
,
#1

Do you currently have any models in production, being used for a product, service, or something else important? Do you ever plan to have any? If so, how do you create and manage your production deployment? How do you train your models? How do you serve your models?

3 Likes
#2

Sir I am planning to build a smart bot that talks and guides people towards a more healthy mental physique, I being an engineer start with the very first tools, pen and paper plan out the model, its requirements (software and hardware) and then do the coding part, I plan to soon make it on production. TF helps to a great deal in doing complex coding in just few lines.

1 Like
#3

Kind of an old post, but I like the idea, so let me try to give it some love :slight_smile:
Yes, at my company I am working on bringing a fire-prediction ML solution into production. The prediction service is part of an IoT implementation in a large forestry area. We gather on the ground data with IoT sensors and collect the data. The prediction service runs as its own docker container which retrieves data from the Prometheus DB and runs it through a classifier and regressor. The latter two are trained by the training container, once every few months with updated data.

The results are shown to the customer (government forestry org.) by means of Grafana Dashboard.
All in all, training the models was 20% of the work, engineering it into production took way more effort.

#4

Thanks TimoKer! I think that the division of the work that you outlined is all too common, especially for the first model that you put into production. I’m assuming that was the case for you?

#5

Yes I’m on my first project on video games creations

#7

Yes indeed you are right. Although I can imagine the production side becomes much more complex for larger scale deployments. So even though one gets better in navigating the deployment side of ML, as the applications get more scale and complexity, so does deployment. Therefore, perhaps the 80% remains a good rough estimate? What do you think, as you have more experience with complex deployments at large scale?

#8

Are you using any ML? If so, care to elaborate? (: