Arnaud C. Email & Phone Number

Paris, Fr

Paris, Île-De-France, France

PRICING MODEL- Improved the existing conversion rate model, (memory usage, computation time, feature engineering, inference, exploration vs exploitation trade-off, prior distribution) to place accurate real time bids (RTB). Training cost improved by 20%.Main python packages: numpy, pandas, scikit-learn, scipyInfra: EC2, S3 Airflow- Researched for bid shading approaches to optimize the cost per action (CPA) or return on ad spend (ROAS). Developed a framework (Python) to benchmark the models offline (on past data). Implemented and desigbed models. Configured A/B tests and analyzed the online results. The best approach improves significantly the CPA by 6%. Data is collected by means of Spark on EMR clusters. Models are trained on EC2 instances with Python. Tasks are scheduled daily by means of Airflow. CI/CD (lint, format & type checks, build, tests) is configured with Github actions.Main python packages: numpy, pandas, scikit-learn, scipy, pysparkInfra: Presto, EMR, EC2, Airflow, github actions, S3, Cassandra- Researched for a multi conversion model (instead of the existing binary one) in order to improve the pricing taking into account the fact the data is censored. Led a team of 3 data scientists. Developed a framework (in Python) to benchmark the investigated models offline. Implemented and designed models. Some of them are trained by bayesian updates. Developed an API (flask) so that the bidder can use our model. A/B test in progress...Main python packages: numpy, pandas, scikit-learn, scipy, tensorflowInfra: EC2, github actions flask, S3CHURN MODEL- Developed a churn model to help clients to identify users about to leave.Main python packages: numpy, pandas, scikit-learn, scipyInfra: Kubernetes, S3, AirflowMANAGEMENT- Managed 2 data scientists.

Palaiseau, Île-De-France, France

Teaching assistant for the following courses:- quadratic optimization- differentiable optimization- statistics

Région De Saint-Denis, France

Senior Consultant Data Scientist in the Big Data & AI Factory of ITNOVEM, a data science consulting subsidiary of SNCF.- Developed a cloud base (AWS) purchase decision making tool for spare part supply chain to reduce the financial freeze cost up to 85%. Identified the most influential variables (hypothesis testings, correlations, mutual information, feature importance). Benchmarked approaches (models: GB, decision tree, NN ; losses: custom objective function under constraints). Implemented the data pipeline with kedro, managed models with mlflow and explained the predictions with shap (Shapeley values).Main python packages: numpy, pandas, scikit-learn, scipy, tensorflow, kedro, mlflow, shap, lightgbmInfra: Databricks, S3, Jenkins- Developed an electric load forecast tool available in a cloud base (Azure) web application in order to place accurate purchase orders. Designed the models as configurable data pipelines. Researched machine learning approaches to improve the electric load predictions (GB, RNN, ARMA, Kalman filter, Gaussian processes).Main python packages: numpy, pandas, scikit-learn, lighgbm, tensorflowInfra: web app, PostgreSQL- Developed a method to distinguish common contact wire wear patterns from singular ones by means of a dimensionality reduction (PCA, NN-PCA) followed by the residual analysis in order to optimize the maintenance.Main python packages: numpy, pandas, scikit-learn, scipy, lightgbmInfra: Databricks, S3- Developed a cloud based (Azure) tool (runs on Databricks) to predict the breakdown probability of freight car over a 4 month period to change from automatic and periodic maintenance to adaptive one. Designed the feature engineering pipeline, the machine learning model (gradient boosting) and explained the predictions (with Shapely values).- Designed a CNN to classify images of rail with and without insulating joints.- Investigated a named entity recognition (NER) solution for incident reports.

Région De Saint-Denis, France

Management of the mechanical modeling and simulation team of the electric traction department of SNCF.- Managed a team of 6 engineers. Responsible for the deliverable validations. Scheduled the work load. In charge of the commercial procedures (client relations, commercial proposals and contracts) and the financial balance.- Conducted the continuous development of the modeling tool of the team, OSCAR, developed in Matlab, to stay aligned with the company strategy. Certified OSCAR conforms to the new version of EN 50318. Modeled the tension variability by analyzing auto tensioning device dynamic data. Proposed a method to reduce the number of rolling stock inline tests by improving the reliability of the modeling and applied it to the TGV 2020.https://www.researchgate.net/publication/303549581_Sensitivity_analysis_of_catenary_geometry_on_current_collection_qualityhttps://www.researchgate.net/publication/350007272_Tension_variability_in_contact_wire_and_messenger_wire_of_auto_tensioned_catenari es- Revised 2 European standards as a member of an European working group for standard revisions: EN 50318 (Current collection systems. Validation of simulation of the dynamic interaction between pantograph and overhead contact line) and EN 50367 (Fixed installations and rolling stock - Criteria to achieve technical compatibility between pantographs and overhead contact line).- Updated the specification of a new overhead contact line geometry contactless measuring system by analyzing the first measures with Matlab.- Involved in a partnership with RTRI, a Japanese Railway Technical Research Institute.

Région De Saint-Denis, France

Mechanical modeling and simulation engineer in the electric traction department of SNCF, the French railway network manager.- Certified several overhead contact lines conform to the Technical Specifications for Interoperability (TSI) Energy by modeling the interaction between them and pantographs by means of OSCAR, a SNCF software dedicated to this task developed in Matlab.- Suggested an overhead contact line renewal prioritization in a French region by analyzing and then ranking the dynamic behaviours of all the catenary designs of the region.

Région De Versailles, France

Designed a few mechanical and electrical systems coupling a drone with a reinforced vehicle. Developed 4 main functionalities: ability to take off, to land, to charge the drone battery up and to hold the drone while the vehicle moves.Patent registering for one of the developed solutions.

Auckland, New Zealand

Analyzed acceleration data collected a precast concret bridge over flexible supports to determine the modal properties and the force distribution.

Argentine

Delivery man in a restaurant.