Tasked with creating an interface for users to request complex ML projects and work alongside Account Managers and Data Scientists to monitor the offline building of models, the accuracy of the models and running the models once they have been delivered.
While many parts of the ML pipeline are automated, you will always get better results when a human is shepherding the work. By fine tuning results and keeping business goals in mind, we assumed we could better serve the end user while providing them with a Zappos or Stitch Fix level of customer service. Allowing for a deep understanding of where their work was in the pipeline and what the current status was.
Not every end user understands what goes into models such as “propensity to buy” so we needed to make the results of the model building step, understandable and worthy of your trust. As through the whole project we would lean heavily on connecting you to an AM that would notate the results in language you could understand.
Once the user was happy with the model, we would need mechanisms to run the model on their data sets, some of which might be incremental or changing as time goes by. We developed the idea of letting users create “automations” that would run based on triggers as well as standard date/time recurring scheduling.
Once a model has been run and the results have been appended to the users initial data set we wanted to visualize the results and entice users to follow up in the various verbs our platform would allow. From viewing the data via data viz to creating deeper segments based on the propensity results, to exporting that data to be leveraged by external marketing orchestration tools.