Raphaël Bournhonesque

• Acted as the principal ML Engineer and technical lead for all AI initiatives, architecting production models while providing hands-on guidance to external contributors on data annotation and training workflows

• Managed data annotation, trained and deployed multiple computer vision models (object detection, image classification) using Ultralytics library, for the Open Food Facts and Open Prices projects. Technologies: YOLOv8, Triton

• [Open Prices] Designed and ran benchmarks to evaluate LLMs in a data extraction task from images, fine-tuned a visual LLM (Qwen3-VL 8B) using Unsloth library. Technologies: vLLM, LLM, Unsloth

• Managed data annotation, trained and deployed machine learning models to extract from images: 1) food ingredients (from OCR, sequence tagging), 2) nutrition data (image + OCR, Document AI). Technologies: XLM-RoBERTa, LayoutLMv3, Triton

• Participated in infrastructure operations, including the configuration of one of our bare-metal server cluster. Technologies:Proxmox, Ansible, zfs, docker

• Participated in the fundraising effort by successfully submitting AI-related proposals to open calls.

• Developed an image similarity service, using EfficientNet architecture for image embedding and Approximate Nearest Neighbors (ANN) for fast image retrieval. Used on a dataset of 2M images. Technologies: triplet loss, EfficientNet, ANN

• Designed and developed a hybrid NLG system to suggest responses to social media reviews. Tags are detected using fine-tuned RoBERTa/CamemBERT models (en/fr language respectively). Technologies: NLG, document classification, transformers library

• Implemented a semantic image search engine. First prototype based on TERAN architecture, final solution based on OpenAI CLIP model. Technologies: Object detection, multi-modal learning, ANN

• Improved our NER engine, by increasing dataset quality and improving NER model. Technologies: NER, CamemBERT, CRF

• Improved the relevance of results returned by the internal search engine (based on Elasticsearch). Decreased user search latency by 60% by developing a dedicated NER service.

• Trained and deployed a Tensorflow NER model (biLSTM+CRF) to detect person names in search queries.

• Designed and implemented an algorithm to detect noun phrases in search queries, using syntactic parsing on the Common Crawl corpus.