Thomas Chaigneau

- Trained multiple LLMs from large to XXL sizes, using Supervised Fine-tuning (sft) and RLHF: Flan-t5, Llama, GPT-J. - Created an optimized the best open-source ASR tool using Openai Whisper and NVIDIA NeMo. - Created an internal framework to efficiently train and expose LLMs on GCP using one or multiple GPUs and request batching when possible. - Maintained a Python SDK, a Slack bot, and a Discord bot that allows users to interact with our Summarization API in seconds.

- Created a Python SDK to communicate with Summarization API, including robust testing and good development practices. - Developed a benchmark tool that automates gathering summaries from multiple APIs. - Updated API documentation and user guides for the different API endpoints.

- Led the training of huge models on multi-GPU instances on GCP and determined the best practices according to the company policies. - Deployed a self-managed MLFlow instance with artifacts storage on S3 buckets, associated with a Python package that handles all user's configuration variables. - Designed Agile/Git workshops to standardize the AI team's project management and code workflows.

- Deployed the first Computer Vision model in production by training a model from scratch, building an API and deploying it on the company private servers. - Initialized deep learning models conversion to ONNX for multiple projects and models by creating guidelines and documentation for my team. - Contributed to develop, document and extend the use of an internal ML app deployment template by supervising releases and upgrades. - Developped multiple applications to showcase ML usage with coworkers by building and deploying Streamlit apps. - Explored a new solution for document classification by fine-tuning a LayoutLMV2 model in order to demonstrate feasibility.