Exa Research Engineer Interview Experience — Two Phone Screens, One 40-Minute t-digest Derivation

Haven't seen anyone share about this company on the forum, so let me be the first to post.

Research Engineer — two phone screens total.

Round 1:
Mostly a PyTorch contrastive loss debugging exercise. Basically just fixing some matmul shape mismatches. Once I got it working, there were no more issues. The interviewer was a young guy, really nice, and he even guided me toward the right direction. According to him, the PyTorch question could have been swapped for a different one related to streaming softmax — pure speculation on my part though.

Round 2:
The original interviewer had a conflict, so they swapped in a Research Manager instead. I checked LinkedIn afterward — very young, Harvard undergrad plus an ML PhD, and had that young-and-arrogant energy. Personally I didn't click with him, and I vaguely sensed some ageism. The question started with discussing how to do percentile estimation, then quickly escalated to having me derive the percentile guarantee with a plus/minus percent error bound, and he wouldn't even let me discuss other non-optimal approaches. I realized afterward that he basically wanted me to derive t-digest live in 40 minutes, which felt like a lot to ask.

Then again, Exa just raised a Series C and became a unicorn, so maybe they've got the capital to be full of themselves...