Over the years we tried multiple management methods to gravitate towards Scrum finally. Having set our standards high, we build great applications for our clients, respecting their time and budget predictions. Scrum is an ideal help in our daily routine. Experimental but systematic, it harmonizes naturally to the cycles of artificial intelligence development. With each cycle (sprint) we carry an experiment to deliver a working model and check if it meets our client's highest expectations.

Here is how a typical AI project looks like:

Understanding the context of your business: (1 sprint)

  • develop an understanding of your business strategy, risks, and goals
  • identify areas that can benefit from machine learning
  • pick the most promising models to implement
  • Outcome: report

Data maintenance: (1 sprint)

  • clean the compose data that might be coming from many different sources and formats,
  • make how to get supporting data (data generation techniques, online sources of similar data)
  • develop baseline models
  • Outcome: dataset and a baseline model

Deep learning: (1-4 sprints)

  • choose a model to use
  • select data augmentation
  • train the model
  • inspect the results to identify next areas of improvments, repeat
  • Outcome: working model with good accuracy

Production Deployment: (1-4 sprint)

  • create an API and/or a web app ( can be done in parallel)
  • export model into the required format
  • plan to retrain the model with updated data
  • Outcome: working solution

Monitoring: (ongoing, not in cycle)

  • develop dashboards to track the performance of your models
  • make a plan to identify and handle mistakes and unexpected consequences