Allaa Boutaleb

Allaa Boutaleb

He/Him

PhD candidate at Sorbonne University's LIP6 lab - Database Team. Husband, gamer, cat-lover and coffee addict.

AI Engineer @ Sujet AI11/23 - PresentParis, FR

As a co-founder and AI Engineer at Sujet AI, I played a pivotal role in establishing the company’s technical vision and leading the development of its core open-source contributions. Sujet AI is a Paris-based startup dedicated to democratizing investment opportunities by creating powerful, transparent, and accessible AI tools for financial analysis.

Mission and My Role

The mission was to break down the barriers in financial analysis by providing open-source foundational models and high-quality datasets, enabling anyone from individual investors to large firms to leverage state-of-the-art AI. My work focused on two primary pillars: building the foundational technology and designing its practical application.

Foundational Model and Dataset Development

A significant part of my role involved overseeing the entire lifecycle of our open-source models. This began with data curation and creation, where I led the effort to build several large-scale, specialized datasets for the financial domain, including:

  • Sujet-Financial-RAG-EN/FR: Over 130,000 question-context pairs in English and French for training retrieval-augmented generation systems.
  • Sujet-Finance-QA-Vision-100k: Nearly 10,000 image-based question-answer pairs from financial documents to train Vision-Language Models (VLMs).
  • Sujet-Finance-Instruct-177k: A massive dataset of 178,000 instruction-following examples for fine-tuning LLMs on various financial tasks.

Using these datasets, I trained, fine-tuned, and released a suite of models on Hugging Face, including:

  • Marsilia Embeddings: High-performance sentence-embedding models for finance in both English and French.
  • Lutece-Vision: A VLM fine-tuned on financial documents for tasks like chart analysis and visual question answering.
  • Sujet-Finance-8B: A fine-tuned Llama 3 model tailored for financial text generation, classification, and analysis.

Multi-Agent Financial Analysis System

Beyond foundational models, I designed and engineered a multi-agent interface to serve as an “AI Copilot” for investment analysts. This system leverages our in-house models to perform complex tasks, such as conducting due diligence, gathering insights from dense financial reports (e.g., 10-Ks), and generating custom summaries. The multi-agent architecture allows for a sophisticated division of labor, where different AI agents can collaborate to analyze information, cross-reference data points, and synthesize findings into a coherent report.

Current Status

As of today, my day-to-day involvement in Sujet AI has been reduced as I now focus on my PhD research obligations at Sorbonne University.

Tools and Technologies

  • Programming Language: Python
  • Core Libraries: PyTorch, Hugging Face (Transformers, Datasets, Accelerate), LangChain, LlamaIndex
  • Models: LLMs (Llama 3), VLMs (Florence-2), Sentence Transformers (SBERT)
  • Infrastructure: Hugging Face Hub (for model and dataset hosting)