Jean-Baptiste Lemoine

Real time automatic car accident report generation from audio transcription. • Built data extraction model based on LLM tailored for car accident report (guided decoding). • Created synthetic data creation pipeline and custom score metric. • Fine tuned LLM (Llama 3.1 8B) on json diff task. • Sub-second latency inference (TensorRT).

Helped Acolad develop AI solutions for Machine Translation, Automatic Post-Editing and Quality Evaluation. • LLM prompt engineering and RAG for Machine Translation Post-Editing, reducing manual effort by 40%. • Train non-technical teams to use AI tools via the Orq AI collaboration platform. • Built serverless GPU-powered backend on Modal for Automatic Post-Editing and Quality Evaluation simplifying exiting Kubernetes infrastructure.

Aive is an AI-assisted video editing platform, it automates post-production, intelligently adapting videos across all formats. • Prompt engineering and fine-tuned LLMs (LoRA) for transcript topic segmentation: +44% accuracy over SOTA. • Developed Vision Transformer model for video scene detection: +20% AP • Built models for black bar detection (Numba-optimized) and color palette analysis (k-means). • Trained image classifiers for video place/scale recognition. • Migrated OCR from PaddlePaddle to PyTorch for 10x faster inference and improved text tracking across video frames. • Transitioned cropping to ML-based saliency detection. • Built optimized video decoder and visualisation tool backed by PyAV and FFmpeg. Profile and optimize model for inference (Nvidia Nsight, mixed precision, TensorRT). • Mentored team with reproducible research tools (PyTorch Lightning, W&B, Ray).

Analytics for TV and radio advertising campaign. I lead the R&D and deployment of Machine Learning models for: • Measuring global incremental impact of TV/ radio ads on advertiser web/app traffic. • Attributing website visitors to a specific ad