Florent Lefort

Project 1:

Creation of a centralized Offers Hub and a pilot dashboard to track offer dissemination activity

Context:

The Offers and Employer Branding department needs a unified, reliable, and actionable view of job offers. The goal is to reduce the time spent manually cross-referencing data, accelerate access to key indicators, and enable product teams to monitor dissemination activity daily.

Results:

- A centralized Offers Hub was made available, used as a common repository by Product Managers, eliminating complex SQL queries and reducing time spent on data extraction.

- Daily activity monitoring is now possible thanks to the automated dashboard, which detects dissemination anomalies in real-time and answers recurring business questions (volumetry, offer attractiveness).

Technical Stack:

HiveQL, Python (pandas, streamlit), Dataiku, Bash, Git/GitLab

Project 2:

Creation of a real-time monitoring dashboard for the usage and consumption of LLM models

Context:

With the multiplication of generative AI use cases within France Travail, the Agency Data Services department must control costs, anticipate consumption drifts, and secure the scaling of LLM models. The objective is to detect abnormal usage and define technical safeguards to prevent overruns.

Results:

- Establishment of per-minute quotas based on statistical analysis of consumption distributions, preventing overconsumption and allowing for quarterly forecast budget allocation.

- Pilot dashboard used daily by Product Managers to track LLM model adoption.

Technical Stack:

HiveQL, Qlik, Kubernetes (CronJob), Bash, Git/GitLab

Project 1:

Creation of a prediction and explainability model to automate the validation or rejection of banking operations

Context:

BPCE advisors must process hundreds of banking operations daily (large transfers, check deposits), manually deciding on their validation or rejection. The goal is to help prioritize and secure decision-making.

Results:

- Deployment in production of a decision-support system used daily by advisors, providing instant recommendations (validation/rejection) for each banking operation.

- Continuous monitoring of the model via a dashboard with an automatic alert system to detect any performance drift and ensure long-term reliability.

Technical Stack:

SQL, Python (pandas, scikit-learn, xgboost, mlflow, optuna, shap, dvc), GCP, BigQuery, Power BI, Teradata, Bash, Git/Bitbucket

Project 2:

Development of a predictive model to identify customers at risk of terminating insurance products

Context:

The termination of insurance products represents a direct loss of revenue for BPCE. Retention actions are often triggered too late or too broadly. The objective is to move from a reactive stance to a prevention strategy by identifying weak signals of customer departure for insurance products.

Results:

- Predictive model deployed in production, identifying customers at high risk of termination 3 months in advance, giving advisors sufficient time to intervene.

- Targeted retention campaigns orchestrated based on predictions, allowing advisors to proactively offer more suitable products to at-risk customers and improve retention rates.

Technical Stack:

SQL, Python (pyspark, scikit-learn, xgboost, mlflow, optuna, dvc), Teradata, Hadoop/HDFS, Bash, Git/Bitbucket

Project:

Development of a matching model to identify the best consultant profiles for calls for tender

Context:

Valeuriad's sales representatives spend significant time manually searching for profiles suitable for calls for tender, with the risk of missing relevant skills. The objective is to leverage the wealth of skills profiles to accelerate pre-sales staffing and identify a similarly skilled employee during mission replacements.

Results:

- Major time savings for sales representatives, who can identify the "Top 5" relevant experts in seconds instead of hours of manual search.

- 2D visual mapping of the skills of over 150 employees, providing sales representatives with an instant view of available profiles and facilitating the identification of replacement candidates.

Technical Stack:

Python (pandas, spacy, sentence-transformers, scikit-learn, flask), Docker, Git/GitLab