I applied online. I interviewed at OneThree Biotech (New York, NY) in Sep 2020
Interview
A standard phone screen assessing general personal fit and technical background, followed by a take-home that was supposed to take 4 hours but took me at least 30 and a subsequent rejection.
Interview questions [1]
Question 1
Code up a modification to a common classifier following pseudocode in a paper, without using sci-kit learn.
First there was a 30 minute Zoom covering my background, a few short technical ML questions and time for me to ask questions;
Within a week I was asked to complete a technical challenge that was supposed to take 4 hours. It took me two weeks!
Interview questions [1]
Question 1
take home problem that asked to code a standard ML algorithm from scratch in python (no scikit-learn) and then modify it according to a research paper that only had pseudo-code
Interview with founder first, and then a data challenge. The data challenge they ask you to code a random forest from scratch for no apparent reason. The paper they give you to implement doesn't actually change the internal structure nor require a rewrite for any reason. I strongly believe in not re-inventing the wheel not necessary. Also, there was an aire of we are smarter than you saying they could solve the problem and write an entire random forest class in 3 hours. I would love to see the scikit-learns decision tree and rf forest written in three hours, even by the main contributor. Then, you had to also write the additions from the paper which were pretty simple and produce standard metrics for the output. The additions of the paper external to the RF dealing with imbalanced classes and metrics were easy. That could be done in a few hours no problem. Anyways, that part just left a bad taste in my mouth. The founder did seem nice and motivated, so that was refreshing.
Interview questions [1]
Question 1
Read a paper and code an rf with the slight modifications for imbalanced classes from a paper. Then return back some metrics after running on the dataset provided. Although the paper doesn't actually modify the structure of a RF, just adds external mechanisms based on k-nn to combine to RFS to help in reducing bias due to imbalance classes. There ask didn't make sense to me and seemed a waste of time.