The first one asked me to implement a Naive-Bayes Classifier on Map Reduce from scratch (ie. no Naive-Bayes libraries), gave me real customer data & I was supposed to run the classifier & pick out the list of customers who should be sent emails based on the classifier's results. All in the course of 1 hour!
The second startup gave me a time-series signal and I was asked to figure out whenever the washing machine was running. When you have electric signals, you can decompose them into fourier harmonics, & running the washing machine leaves a distinct signature - you'd get extra harmonics if you ran the washer.
The third one asked me to flag a bunch of potential delinquent borrowers from a 30-year home loan repayment schedule. Risk prediction
All 3 were fairly code-heavy (150-200 scala loc)