Gregory Vial

An implementation of the Relaxed Lasso paper by N. Meinshausen (2006).

I implemented the paper in Python in a sci-kit-learn-compliant fashion.

Relaxed Lasso is an improvement over the standard Lasso regularizer for linear regression. It lets you control the number of variables retained, and the amount of regularization applied using two separate hyperparameters, leading to sparser models rather than classical Lasso while achieving equal or lower errors.

The idea is to make personalized predictions based on each driver's habits, depending on inputs such as day of the week, time of the day, and current location.

I was the sole inventor of the method, and I also developed an implementation that used Python and TensorFlow.
The model was then made accessible through an API (Flask based), wrapped into a Docker container, and exposed through Azure Web App. It later served a fleet of approximately 50 demo vehicles.

The project led to a patent filing; see attached link.

An end-to-end system makes it possible for everyone to track their home water consumption on their mobile. The system comprises a PLC-based solution installed at home, connected to WiFi, and sends data through an API to the Azure cloud. Data is then stored on a Postgres database and served to consumers through a web-based front end.

Using evolutionary strategy, I optimized the design of an electric machine included in the car powertrain.
Based on an existing machine design, a set of rules that tie together all degrees of freedom, and expert software that computes several target features for a design, I used several evolutionary strategies to explore the parameter space and come up with a design that beats the human experts, hence creating superior electric machines.

The purpose of the project was to divert the standard use of the knock sensor (a cheap vibration sensor used to detect auto-ignition in an internal combustion engine) and identify if it could detect defects in a car.

The first phase of the project consisted of data capture. We equipped a car with multiple sensors and then performed various drives with defects applied to the car, like low tire pressure, parallelism issues, etc.

I then analyzed the vibration data of each run and compared a standard car (no defect) with the various defects applied.
Using multiple machine learning techniques, I developed a data pipeline that included data pre-processing and a machine learning model that detected various defaults with high accuracy.

The project led to patent filing; see the link.