I joined the French startup AnotherBrain to take part into their first product in quality control for the industry. We develop a light capable of making photos which highlight the defects to inspect on industrial products. This projects brings a lot of innovation based on computer vision (OpenCV and deep learning) and learns the normal distribution of objects to deduce the abnormal pieces or the defects.After a quick introduction within the R&D team dedicated to pure research, I decided to join the R&D team specialising in AI solutions for the product. I develop the last algorithms, improve their performance, customise them so that hey meet the clients' needs (real time, light-condition invariance, defect size to detect). I am responsible for following the progress for three major clients of the startup. I supervise the R&D (taking myself part to it), I assess the performance of the solution and I make sure all the teams (embedding, mechatronics, sales) are aligned, in relation with the product manager. This mission requires me to implement / adapt Python code (PyTorch and OpenCV to name a few), to design G-Streamer pipelines and adapt C++ / CUDA plugins. I was able to improve the quality and efficiency of the machine-learning code by introducing visualisation, inference metrics and CI.

Within a team of 20 data-enthusiasts, I took part to the various projects of Sopra Steria. These projects covered a wide range of sectors including defence, road safety, public finance, employment and I was leading the R&D strategy of the team.In each project, I was responsible for enhancing the data of the client (eventually bringing external data into play), in order to address his problematics or to raise new opportunities, in association with the consulting teams. In a constant care of guidance in his digital transformation, we proposed innovative solutions with a pedagogic and scientific approach (definition of metrics of the developed algorithms) with the final objective of industrialisation. I list some of the projects I led: tool for predictive maintenance (Machine Learning with Scikit-Learn), module for car make and model recognition on photographs (Deep Learning, PyTorch), hotel-pricing retrieval (web scrapping from booking websites, statistical analysis), data audit, automatic car-plate reading (Deep Learning, Keras and dedicated libraries), invoice reading (Tesseract), anomaly detection in tax declarations (Machine Learning with Scikit-Learn), NLP (Deep Learning, Keras), AI strategy definition and R&D (automatic data generation using GAN and VAE). I also took part to the formalisation of the Datalab's strategy (recruiting, staffing, skill grids and training).

I did my PhD thesis with the CEA Saclay (DSM/Irfu/SPP) within the ATLAS collaboration (3000 researchers). I was based at CERN (Geneva) during three years at the heart of the experiment, in an international professional environment. My position offered me a great visibility in the collaboration and I could defend my work during the internal meetings or conferences (posters, talks). Writing my thesis and articles has been the occasion of summarising my searches in a constant effort of vulgarisation and accuracy. I was responsible for the optimisation of the muon isolation (the muon is a heavier particle as compared to the electron). A particle is considered well isolated if it is far enough to any other particles in the detectors. From the energy deposits, I reconstructed the particle trajectories (clustering on images and regression of the trajectories). From these I could deduce the isolation. This calculation was of major importance to the collaboration and I implemented the tools (C++ packages and Python scripts) accessible to all the researchers in production.In the mean time, I took part to the analysis of the Higgs boson decaying into four leptons (muons and electrons). I helped for the production of the samples (files from simulated and real data). I validated and optimised the selection criteria of the events for the analysis (rare-event analysis similar to anomaly detection in machine learning). My work led to some publications. Here are the most relevant references:ATLAS Collaboration, Aad, G., Abbott, B. et al. Eur. Phys. J. C (2016) 76: 292. ATLAS collaboration, Aaboud, M., Aad, G. et al. J. High Energ. Phys. (2017) 2017: 132. Lesage, A. A. J., Study of the fiducial cross section of the Higgs boson decaying into four leptons and study of muon isolation in the ATLAS experiment at the LHC.

On top of physics and computing lectures, I worked within the Semiconductor physics group of the Cavendish Laboratory, where I studied the properties of the one-dimensional electron gas (quantum wires). The objective of my research project was to analyse an anomalous observed in the conductance of these wires.In order to do so, I used a multiplexing technique to measure 256 devices in a single cool-down (cryogenics), enabling the measurement process to be automised as well as to increase the available statistics. Because of the large amount of collected data, I implemented a routine for automating the result extraction which allowed to extract local minima on images obtained from the conductance measurements (3D maps). This routine accelerated the processing of the whole team.My work led to the following publications:- Smith et al. Phys. Rev. B 90, 045426.- Lesage et al. J. Appl. Phys. 117, 015704 (2015).- Smith et al. Phys. Rev. B 91, 235402.- Al-Taie et al. J. Appl. Phys. 118, 075703 (2015).- Smith et al. Phys. Rev. Applied 5, 044015.