Interviewed 16 October 2018. Lightly edited.
If you’re interested in being profiled, drop me a line! I’m interested in speaking to early-mid career data scientists (hence, n=1 year).
What do you do?
I’m a data scientist at KeyMe.
Our key value prop is in the tech, e.g. deep learning, which we use to scan keys and model what the output should be.
What did you study in school?
I studied mechanical engineering and physics at Penn.
What made you want to do data science?
On suggestion of a friend, I applied to data science internship, which was code heavy but seemed more interesting than software. The problems were open-ended and cool: what’s the problem, how can we solve the problem, flexible on which techniques to use to solve the problem.
What did you expect to do as a data scientist?
- looking for trends
- correlations that others weren’t finding
- predictive modeling
- a lot of communicating results to management and arguing for a path.
What do you actually do?
At first, I was the second hire on the team, with various transitions, that meant a lot of ETL / pipeline firefighting and small asks at first.
Now, it’s more more like independent research, into some optimization models, but not a lot of ML. Another big element is data evangelism, trying to bring people up to speed with data / share data more broadly. An example is the marketing analytics team:
- teaching SQL and aggregating data in a given way to help them
- walking them through some data analysis questions
- recognizing common pitfalls (e.g. causation vs correlation)
What tools do you use?
Python, TensorFlow and Keras, Pandas/Numpy, FB Prophet, Redshift / SQL, Luigi for ETL, Tableau for visualization, Excel for everything which goes into PowerPoint
Which teams or roles do you work the most with?
Analytics is centralized - we work with marketing (campaign performance), engineering (data integrity, infrastructure), as well as ops (supply chain optimization, scheduling).
Information flows through analytics and it’s hard to scale that to all teams - no single source of truth, need to define ground truth metrics and statistics, making sure everyone can make good decisions. A centralized approach is good in the short run.
How did you have impact on the organization?
Lot of different areas:
- Revamped store selection model from third-party to in-house model.
- Utilized new data sources and engineered new model which is 2x as good at predicting store usage.
- Inventory utilization: each machine can only have a certain number of slots for key molds, which ones will have the most usage. We made this 50% better.
- Refund rates are down drastically because of improvements in error rate, by understanding the purchase and refund funnel.
- Dynamic pricing - increased core revenue by 25%.
How do you do experimentation?
We typically set up user segmentation and assign users to a given experiment. This requires coordination with our front end engineers and product team.
What’s the most important thing you learned?
It doesn’t matter what you find unless you convince others to take action on your results.
Any other advice you have?
The difficulty of data science code is overhyped: concepts are more difficult than most practice. In the real world, cutting-edge ML isn’t that important, it’s not Kaggle where we’re going from 95→99%. It’s more about problem solving and ability to independently learn and research. most peoples’ problems are not 100% unique, you can find research that’s gets you 80% of the way there, your value is in knowing what to do and how to apply the existing research.