Thomas Portier

Spreadauto automates financial data processing by detecting and organizing financial tables using NLP. Key improvements include:

Data Extraction: Identifies and extracts financial tables from reports.

Code Upgrade: Migrating to a newer Python version for better performance.

Model Enhancement: Retraining models for improved accuracy.

Automation & Scalability: Using Docker & Kubernetes for efficient deployment.

This project automates financial data extraction and improves scalability. Key components include:

High-Performance System: Runs on 4 Nvidia A100 GPUs inside a Docker container.

Data Extraction: Uses LLM with RAG to analyze financial reports.

API & Database: REST API for data access and storage.

Automation & Scaling: Managed with Jenkins, Ansible, and Kubernetes for seamless operation.

Both projects aim to streamline financial data handling, reduce manual effort, and improve accuracy.

Creation of a complete AI solutions in production:

- Extract text from contracts (OCR)

- Classify contract clauses

- Identify and extract key terms

- Link related words (e.g., address + company name)

- Improve accuracy with active learning

- Develop an annotation tool

- 80% accuracy with 50 contracts, 90% with 100

- Stable and scalable with optimisation on RAM & CPU usage

The mission of this project is to create an AI-powered system that can read and understand legal contracts just like a human would.

At its core, the system follows a structured approach called Logical Question Answering (LQA), where it extracts key information from contracts by answering predefined questions. For example, if asked, "Who are the parties involved?", the system would identify and extract the correct names.

To achieve this, we use advanced AI techniques such as image processing, OCR (text recognition), and language analysis. The system can recognize legal clauses, detect differences between documents, and even learn from user feedback to improve over time.

We also developed tools to make this process more efficient, including an annotation application for reviewing extractions and automation tools for model updates and performance tracking.

This entire solution is built as a scalable SaaS (Software as a Service) platform, managed by a development team in Shenzhen, and continuously improved based on customer needs.

Stock Exchange Prediction Optimization Project

Objective: Optimize real-time stock market prediction by extrapolating stock option data.

Key Subprojects:

Developed a Python web scraping tool using Selenium & BeautifulSoup.

Scrapes stock option data in real time.

Stores data on a web server using a Redis database.

Applied deep learning & machine learning models:

LSTM, Perceptron, Random Forest.

Real-time peak detection for risk assessment.

Speed Optimization:

Algorithms converted to C++ & Fortran for processing under 1 second.

Used OpenCL & Vispy for real-time data display.

Achieved latency under 30ms.

This project is designed to predict stock market trends more accurately and in real time. The goal is to help investors make better decisions by analyzing stock option data as it changes.

Here’s how it works:

Collecting Data: A smart program automatically gathers stock market data from the web in real time and stores it securely.

Making Predictions: Advanced AI models analyze the data to detect patterns and predict market movements, balancing accuracy with risk.

Speed Optimization: The system is designed to process predictions in less than one second, making it extremely fast.

Real-Time Visualization: The results are displayed instantly with cutting-edge graphics, ensuring updates happen in less than 30 milliseconds.

In short, this project combines AI, automation, and high-speed computing to create a powerful, real-time stock prediction tool.