Ismael EL ATIFI

Working on ad targeting using machine learning. Built a Python library (with CI/CD) to train LightGBM models for user segments prediction. Developed a POC for doing the same with a single deep learning Pytorch model (instead of multiple LightGBM models).

Python development and GitHub CI/CD. I mainly contributed on FastAPI backends including one for models tracking using PostgreSQL database and MLFlow Model Registry. I was in a scrum team of 3 devs.

Python development around ML based ads recommendation. - Implemented a backend with FastAPI responsible for model deployment with Databricks Model Serving, and allowing multiple models to be served in production at the same time. Models were tracked using MLFlow and the Docker container of the Fast API app was deployed on Kubernetes pods thanks to ArgoCD. - developed the A/B testing dashboard with Streamlit to compare the Click Through Rate of multiple models, number of displays and other metrics. - implemented PySpark jobs with both SQL and dataframe API

Doing software development in a Scrum team of 4 devs. We develop and maintain Python backends to create and stream business objects by watching for changes in PostgreSQL databases. All our code runs on AWS (Lambda, ECS, Step functions, SQS, Glue PySpark, DynamoDB, PostgreSQL, Kinesis, DMS). We use Gitlab for CI/CD.

Worked on optimization of promotional calendars (Python codebase in production). Future volumes sold for thousands of candidate promotional events are predicted with Gradient Boosted Trees (XGBoost), and then Mixed Integer optimization (PyOMO/Gurobi) is used to find the optimal subset of events which maximizes ROI given business constraints.

Worked on multimodal (image+text) document classification and information extraction from scanned images and smartphone photos. I built an information extraction Deep Learning model, with OpenCV and PyTorch, which performed really well on scanned documents (based on CNN + Transformer).

I built Computer Vision Deep Learning models from scratch to analyze photos of ID documents (ID cards, passports and driver licences, for many countries). The tasks I worked on were classification, localization, text detection, OCR and font recognition for anti-fraud. I coded in Python and used libs like OpenCV, imgaug, Tensorflow and PyTorch to train neural networks and Tensorboard to follow model training. I achieved very good results with synthetic training data and extensive data augmentation. I often customized the neural networks, loss functions and metrics to maximize predictive performance. I also implemented in Python a template alignment algorithm (based on RANSAC) to improve our text detection and OCR solutions.

I first worked about 2 years as a C++ backend engineer on reservation systems with high TPS and using a big Couchbase cache to store flights availabilities. Then I joined the Research Optimization and Innovation team as a Data Scientist to work on improving Amadeus flight search product using Machine Learning. I built a gradient boosted trees model to better rank travel solutions. Main tools I used are : - Python - Pandas - PySpark - Jupyter Notebooks - Elastic Search - H2O/CatBoost libraries for decision trees

Main missions with the Algorithm team at Movea : - Design, development in C++ and test of algorithms for gesture recognition (Machine Learning) based on sensors data (accelerometers and gyrometers). - Developed a software solution to precisely measure movement using images from a camera and the Computer Vision library OpenCV (C++). - Developed multidimensional analysis functions (Principal Components Analysis and Linear Discriminant Analysis) using C++ linear algebra library Eigen. - End-of-study project (7 months) : Development in Java of an Android application using the Google Activity recognition API to compare results with our activity recognition solution. Designed and coded in C++ an algorithm to recognize “digit” gestures made with a special remote. Good classification rate was above 90%.