Research on audio source separation Deep Learning algorithms and state-of-the-art. Audio source separation is a vast field of research in Deep Learning with multiple uses. I worked on this subject and did all the research as if I was to lead an audio source separation project. No coding was involved, the objective was purely to research, benchmark, look at what had already been done, break down the different algorithms used and assess what made one better than another, then write a paper about all the collected information. This covers spectrogram methods and waveform methods, such as U-NET.

Machine Learning model capable of identifying brain tumor location based on medical imagery. The model is a SCV pixel classifier, as opposed to an image classifier. Several optimizations were possible, such as using a multi-layer perceptron or random forests, or simply using more context for classification.

Context: PrestaCop, a company specialized in service delivery for police forces, wants to create a drone service to help police systems collect parking tickets. A camera with a pattern recognition software identifies license plates and characterizes infractions.
The drone simulator is written in Go and can simulate multiple drones sending messages regularly. All messages are sent on a Kafka stream and depending on whether the drone requires assistance or not, on one topic or another. Azure EventHub stores all messages in Kafka stream into an Azure DataLake, and for messages on the assistance topic, sends an email indicating what the problem is. We also use DataBricks (code in Scala) to analyze the stored data and obtain interesting statistics. The whole cloud part is deployed using IAC.

A Deep Learning model that takes a captcha image as the input and returns its predicted text. The code is written in Python and hosted in AzureML, which allows us to, using Keras and Tensorflow, train the model, save it and export it. This is so that the model can basically be loaded and used as an API using the Azure Machine Learning studio, which we then feed an image (preprocessed), and it returns the captcha text prediction. The model uses a CNN to extract the features and predict the text from the image.