Data Scientist
SensCritique · Paris
Design and production deployment of AI systems at the scale of the SensCritique catalogue (785,000 works, 35 million pieces of content, 109 million interactions): recommendation, spam detection, LLMs and AI tooling for business teams.
- Hybrid recommendation engine: designed and shipped a catalogue-scale engine (ALS and vector retrieval, LightGBM reranking) over 785,000 works and 109 million interactions. Result: relevance of suggestions doubled and diversity of recommended works multiplied by 19.
- AI spam detection (pair work): co-developed and deployed a model detecting unwanted content across 35 million pieces of content. Result: over 90% precision on documented cases, with up to 12% of profiles identified as spam.
- LLM & MLOps: fine-tuned LLMs for sentiment analysis (irony, cultural nuance) reaching 94% accuracy and improving the relevance of per-work scores by 15%. Automated retraining and model monitoring with Airflow and MLflow: time-to-market cut from 2 weeks to 2 days.
- Generative AI for business teams: set up an MCP server (team of 3) to query the internal database in natural language, without SQL. Built an LLM editorial-monitoring pipeline with the Communication team: 3 hours of monitoring saved per day and a reaction time under one hour. Sales & Communication dashboards designed in pairs.
