Please excuse this tortured metaphor.
AI / ML solutions are not silver bullets, they're more like tenuous emulsions in cooking. The data used, features selected or engineered, algorithms selected, validation/testing methods, etc, are all ingredients that have to be mixed juuuuust right for your target objective.
Get any of these wrong in the mix and your emulsion breaks like a bad mayonnaise. Don't define the objective of what you're trying to cook and what it should look / taste like and you'll have no idea what you're making. If your data drifts and you're not whisking your models enough, your mayonnaise breaks again. Attempt to apply your model to an environment too different from its origin, and your mayonnaise breaks. Even when you do achieve the texture, consistency, and flavour you're looking for you better have traced every single step so you can reproduce the recipe at scale in a production kitchen.
Your data scientists, ML engineers, etc are all the chefs in your kitchen. Listen to them. There are certainly lessons from SDLC program management that can be applied, but it's important to know that AI /ML is a very different and has subtle chemistry. There will be a million different concerns. From the temperature of the butter to the coarseness of the salt. While these concerns may sound insignificant they can make a huge difference between achieving your objective successfully or eating a repulsive meal of your own making.
Just like in the kitchen there really are no shortcuts or free lunches. There are no cutting corners here, only getting better at cooking and planning your recipes. If you want to succeed in leading a kitchen, you should know a little about cooking yourself and help the chefs of the kitchen reach their full potential by getting them the freshest ingredients, best tools, and continuous training possible. Make sure requirements are solid. This involves working with problem owners to define the actual problem as an ML problem and that performance criterion are defined.
Use project management processes as tools and pick the right tools for the job. Processes might just as well harm performance as help depending on how they are implemented.
Manage expectations. What you do is considered magic by many, and you need to ground it in reality. Since you'll communicate to non-expert stakeholders, executives, etc. this is crucial.
Regarding ways of working, I like short sprints, 1-2 weeks, focusing on iterating fast. Due to being data-driven, ML results may be difficult to predict beforehand (pun not intended).
Remove obstacles as early as possible. I like catching these during daily standup and sorting them out right away if possible.
The project never makes it to production due to failed requirements capture, failed customer understanding, or unexpected manpower needs. As a PM, you can directly impact the latter and often impact the former.
The project never gets off the ground because data isn't procured, either ever or in time, and sometimes due to missed expectations about the required manpower to get the data. Again, as a PM, you can make sure the latter is avoided
The project goes on endlessly without connecting to product needs. Again, you can prevent this.
What gets measured gets managed, what gets managed gets done.
Comentarios