Arnaud Fouchet

In this market, you have to use state-of-the-art machine learning to be competitive, what we did at Dolead.

We ensured that we stayed competitive by regularly doing an A/B test of our bidder against Google's. In 2017 and 2019, we made more margin using our bidder; Google's bidder got more volume, but with a higher cost per acquisition, and overall was less profitable.

To achieve this performance, we used many techniques:
• Deep learning and transfer learning; learning the deep architecture of the neural network on one task where we had more data, then fine-tuning on our pricing task, which had fewer data.
• Mixture of models; using the average of several models got better performance than using the best one. Also, instead of using the average, you can learn your optimal weighted average via machine learning.

A dating app where people don't match on their picture, but by a subject they'd like to talk about.

I built the back end and front end using Python, Django, and React Native.
We had a very experiment-driven process, so we made plenty of focus groups to validate our app, ideas, and marketing campaign.

The application was released in mid-December.