Data Scientist - Freelance
Improving evaluation practices of detection models in production: designing robust test datasets, implementing A/B testing protocols, standardizing KPIs, and developing production-grade code.
About this role
Veesion is an innovative start-up specializing in computer vision for detecting suspicious behaviors in retail environments.
As part of the Data team, my role is to improve the evaluation practices of the detection models used in production. The objective is to strengthen methodological rigor across the testing and validation processes of the algorithms.
A key part of the mission involves defining what "good evaluation" looks like in this context — from dataset design to statistical significance — and then building the tooling to make that evaluation systematic and reproducible.
I also work closely with the product team and engineers to communicate findings and support decision-making around model deployment.
Key contributions
- Designed robust test datasets: defined representative selection criteria and created a metric to assess dataset representativeness.
- Implemented A/B testing protocols to compare model performance prior to deployment, ensuring statistical significance and metric robustness.
- Standardized evaluation practices: defined relevant KPIs per use case, automated benchmark workflows, and formalized validation processes.
- Developed high-quality production-grade code (unit tests, documentation, reproducibility) for long-term maintainability.
- Conducted exploratory analyses and communicated insights to the product team and engineers to support deployment decisions.