It's pretty nebulous, but the work output of a data scientist would be a predictive model, where those results are either useful for some business unit (forecasting), or capable of being shipped as part of a software system and product.
Using Uber as an example, a data scientist would figure out the algorithm/model for assigning the next driver when you request a ride, the model for the shortest route (maybe), and delivering a model to correct GPS works in cities with huge buildings so drivers know exactly where to pick people up when they stand next to skyscrapers.
It requires a ton of infrastructure to do good data science work, since you not only need to validate that the model works using the exact same data in testing/production, but you need to take code that a non-SWE runs, integrate it into the build, figure out the right operational metrics, then deploy as part of some release strategy.
The model is really the smallest part of that process, but occasionally, you can get a huge lift by having someone apply a lot of interesting math.
Using Uber as an example, a data scientist would figure out the algorithm/model for assigning the next driver when you request a ride, the model for the shortest route (maybe), and delivering a model to correct GPS works in cities with huge buildings so drivers know exactly where to pick people up when they stand next to skyscrapers.
It requires a ton of infrastructure to do good data science work, since you not only need to validate that the model works using the exact same data in testing/production, but you need to take code that a non-SWE runs, integrate it into the build, figure out the right operational metrics, then deploy as part of some release strategy.
The model is really the smallest part of that process, but occasionally, you can get a huge lift by having someone apply a lot of interesting math.