Discuss with pathologists and other data scientists in multi divisions for the best strategies to apply AI for histopathology whole slide images.
Build the ETL pipeline for data collection, data curation, and data anonymization.
Cloud infrastructure management.
Deep learning models development on several datasets to detect lesions in WSI

FishCounter project: AI counting system for fishing boat.
Interface with stakeholders, developers, and technical support to determine optimal specifications for the system, both on algorithm and hardware.

Integrate AI to fashion industry to make recommendation and predict trending fashion.
Investigation, design, and implement scalable applications for data identification, analysis, retrieval.
Build image classification model and clothes segmentation model.
Predict trending color on Amazon dataset with Signal Processing algorithm and libraries.

State of the art survey research on event camera, a new type of camera that captures the brightness change in each pixel, which has high dynamic range, high time resolution, redundancy free, and no blur. The output was a report on this camera and current state of the art in the event-based domain.
Prove of concept using the event camera in order to attract investment from stakeholders. The product was a hand gesture recognition system using an event camera and spiking neural network (SNN), and the energy consumption analysis comparing with the frame-based counterpart. From this work, a coworker brought this demo to show to the manager in Japan.
Other tasks involved being an engineer internship advisor, Meetup presentor about Spiking Neural Network. My work was presented at the NeuroSTIC day at Nice Sophia Antipolis University.

Project Flaxscanner, an OCR software to extract scanned document to text to enhance the repetitive work of human. In this project, my task concentrates on the data augmentation to improve the accuracy of the Flaxscanner. My approach was using Variational AutoEncoder to create synthetic character from online hand writing dataset.The pipeline was made, but the quality of the synthetic character was too good compared to real images, so the synthetic data can only be used to create new style of handwriting but can not replace completely the training images. Although the result was not as expected, I learnt a lot from this work, how to implement deeplearning models and many deeplearning techniques.

This thesis takes a role in the VIRTUALIS project , partner with companies like Thalès and other PMEs. The objective of this project is to create an intelligent surveillance system. This thesis contributed 2 methods to detect abnormal events in the building. The data that we used to analyze were capture from the building of Thales in Paris Palaiseau, which were trajectories of people inside the building and other signals (badge, infrared sensor). The first method was a statistical classifier based on multiple metrics such as stay duration, moving speed, etc. The second method was a kernel method to map the trajectories into a high dimentional space, and apply One Class SVM classifier to detect the anomalies.

Award