Forged in fire: Agile project management lessons from the frontline
There are a huge number of factors which determine whether a data science project in a health or care setting is successful, and two key factors are the complexity of the project and the level of operational pressure to meet specific deadlines with the work. This talk will describe the learning journey of an NHS based data science team who found themselves delivering a project that was both complex and with a high degree of operational pressure.
In particular, there were two methdologies that the team adopted which were essential to ensure they were able to deliver a stable and full featured model to a range of customers. The first method was an agile delivery mode which used a project management process inspired by scrum- but not full scrum, which we did not find useful for data science projects. GitHub releases were used to deliver well defined and regular updates to the model code at the end of four week sprints. The second method drew on the principles of MLOps and emphasised reproducibility of data pipelines and outputs as well as continous QA and code review to ensure that the code in production is always efficient, error free, and model results can be regenerated in the future in case of audit or review. This talk will describe the process of developing these processes within the team, the lessons we learned along the way, and future plans.
Forged in fire: Agile project management lessons from the frontline
There are a huge number of factors which determine whether a data science project in a health or care setting is successful, and two key factors are the complexity of the project and the level of operational pressure to meet specific deadlines with the work. This talk will describe the learning journey of an NHS based data science team who found themselves delivering a project that was both complex and with a high degree of operational pressure.
In particular, there were two methdologies that the team adopted which were essential to ensure they were able to deliver a stable and full featured model to a range of customers. The first method was an agile delivery mode which used a project management process inspired by scrum- but not full scrum, which we did not find useful for data science projects. GitHub releases were used to deliver well defined and regular updates to the model code at the end of four week sprints. The second method drew on the principles of MLOps and emphasised reproducibility of data pipelines and outputs as well as continous QA and code review to ensure that the code in production is always efficient, error free, and model results can be regenerated in the future in case of audit or review. This talk will describe the process of developing these processes within the team, the lessons we learned along the way, and future plans.