Thomas Portier

Spreadauto project: Use of NLP alogirthm to detect, extract and parse financial tables. Code base refactorisation with updates from old to new python version. Retraining models and build its automatic workflow. Creation of a docker container to be used with kubernetes. Valuation Report project: Setup an LLM with VLLM on docker along its deployment with jenkins and ansible with queuing capabilities on a server with 4 A100 Nvidia GPU. Setup in production an LLM with RAG technologies to extract financial data from financial reports. Creation of a REST API along a DB persistency, an MlOps and Devops pipeline using Jenkins and kubernetes for production.

Every contract build inside Gino workflow has a skeleton known as Logical Question Answering (LQA). This LQA is composed with several question-answers items such as: Question: What is the parties ? Answer: "GinoLegalTech" The main task was to obtain this LQA from any external contract using nlp skills. Such task has been done using image processing, OCR, word embedding, classification, NER and depency parser. Extraction are various and can be added by anyone. Link between such extraction are supported. Workflow from user experience are trained to improve all the models. Other task such as clause detection which need proofreading and differences between two texts has also be done. Making annotation application tools in Qt to optimize it. Creating some tools for MLops and Devops optimization (Docker, Automatic models updates with their metrics, etc ...) Collaborating with cross-functional teams and customers to identify product improvements. Managing and implementing the project in a SaaS solution with a developers team based in Shenzhen

This project has for objective to optimize prediction of stock exchange. The main idea consisted in extrapolating stock share option data in real time. It is separated in several distinct sub project. 1. Obtaining a data set: to obtain such data, a web scrapping python program has to be developed. It uses selenium and beautifulsoup to parse html data. It must also be in real time. As soon as the web scrapping application took the stock share option information, it is stocked on a web server in a redis database. 2. Use deep and machine learning algorithm to extrapolate those data: LSTM, percrepton, random forest and a real time peak detection has been used to take the best decision compared to the risk. 3. Optimization - speed calculation: every algorithm has been optimized to process the result with less than 1 second thanks to C++ and fortran algorithm conversion. 4. Visual display: tools such as OpenCL and Vispy has been used to display such result with a latency of less than 30 ms.

• Tests and parameterizations of Sismage Neural Network to model seismic attribute from substacks, tutorial proposal. • Several Java plugins for Sismage's software were created

• Magnitude and Hypocenter calculation • Seismic catalogue creation on Zmap

• Measurement on field and data processing on MatLab