Arthur Gold

Until now, humans were responsible for deciding in less than 15 seconds whether a transaction should pass or be escalated to an agent for investigation. I contributed to the suspicious transaction detection algorithm and enabled the production deployment of this project in a highly secure environment.

My contributions were as follows:

- Improved algorithm performance by adding a data pre-processing layer and feature engineering.

- Created a pipeline on VS Code from notebook drafts and implemented proper code standards. (unit tests, integration tests, type checking, code formatting, creation of a .pylintrc to define adopted code standards...)

- Added tooling on VSCode to unify "best practices"

- Optimized Python code to reduce execution time (using pandas parallel, among others)

- Implemented a Gitlab with a clear Merge Request process within the team that respects certain requirements and integration into Jira.

- Managed the team's Jira

- Implemented the CI/CD pipeline for the deployment of a critical application.

- Implemented the model versioning part

- Integrated the following tools into the CI/CD pipeline (Sonarqube and fortify for static code analysis and test coverage, and Nexus IQ for dependency analysis).

- Created tests for the application (unit and integration).

- Assisted in creating the appropriate architecture for the project.

- Requalified the most relevant KPIs for the business.

- Monitored model performance.

- Implemented monitoring in Streamlit (data drift and model drift)

- Bash scripts for the pipeline. (pre-commit, pre-push)

- Bash scripts for automating data transformation.

- Deployed a Dockerized API for the model.

- Deployed an explainability application (LIME, SHAP, ICE) for the machine learning model.

POC to develop two static web apps for Pernod-Ricard:

- An application where a Pernod-Ricard Marketing France team uploads PDF files and receives an Excel report with all relevant extracted information.

- Another application where another Marketing team sends scanned invoices with sometimes poor document quality. I developed a high-performance OCR (Optical Character Recognition) capable of detecting all fields of interest and a simple user interface to involve human intervention in exceptional cases where the algorithm has doubts.

During my internship, I worked on evaluating over 100 neural network architectures on GCP for reinforcement learning-based quadcopter control.