cyril PETRIS

Development of an intelligent multi-agent system to automate the creation of ETL processes towards the OMOP Common Data Model (CDM) standard. This project aims to drastically reduce the development time of ETLs, a process that typically takes several weeks. Design of a collaborative multi-agent system (Deepseek-R1, Gemini): Architect Agent responsible for analysis and validation: interprets business logic from Rabbit in a Hat specifications, detects basic inconsistencies, and structures a code-oriented action plan. Code Agent that executes the plan: generates the SQL code for the ETL and a complete set of unit tests to ensure that the business logic and technical constraints are correctly implemented. Validation on Reference Data: Full implementation of the ETL mapping specifications on well-known healthcare datasets, MIMIC-IV and Synthea, serving as a foundation for validating the multi-agent workflow.

Contributed to enhancing a statistical interpretation solution for rapid diagnostic tests. The goal was to extend an existing algorithm, initially limited to interpreting tests with a maximum of two to three result lines, to create a more versatile and fully automated system. Algorithm Evolution: Scaled the statistical algorithm's interpretation capability from 2 to 3 result lines (in addition to the control line), making the solution compatible with more complex multi-panel tests. Automatic Test Classification: Developed a computer vision model (based on MobileNet) to automatically identify the type of inserted test, achieving 89% accuracy. This model allows for the correct version of the analysis algorithm to be applied. Image Segmentation with YOLOv8: Trained a YOLOv8 model to precisely extract the result window from the test image. This module reached 82% segmentation accuracy, ensuring that only the relevant area is analyzed. Embedded System Deployment: Integrated and optimized the entire workflow on a Raspberry Pi 4 using Docker containers, resulting in a total inference time of just 1.5 seconds per test.