Matthieu
Olekhnovitch

Description

As a Quant Researcher Intern at Société Générale, I worked on mid-frequency trading strategies, specifically on the development of predictive models for volatility and predictability. My main focus was on recent academic approaches and implementing them in a real-world trading environment. Topics included: volatility prediction, predictability, factor models, and more.

Technologies Used

• Pandas
• PyTorch

Supervisors

• Dounia Ouacha
• Selmen Jnifen

Description

Founded laive.ai, an API-first platform that empowers developers to create sophisticated AI agents with advanced knowledge processing abilities. As technical co-founder, I architected a comprehensive ML infrastructure that provides developers with a full stack of endpoints for document understanding, contextual retrieval, and intelligent task execution. Built core features including real-time semantic extraction with both dense and sparse vector search, automated knowledge graph construction from diverse data sources, and context-aware information synthesis. Engineered a highly scalable architecture able to process 100K+ API calls daily, with dedicated ML pipelines for continuous model improvement. The platform supports various agent frameworks and LLM integrations (Python SDK, Langchain, LlamaIndex, etc.). Implemented authentication, rate limiting, and monitoring systems while maintaining availability. Set up Devops and CI/CD pipelines.

Technologies Used

• System Design
• NLP
• Retrieval Augmentation
• DevOps

Description

Conducted cutting-edge research in automated data augmentation for time series analysis at the University of Luxembourg's Security and Trust research unit. My work focused on extending the AutoAugment algorithm to handle complex, high-dimensional time series data across various domains including medical diagnostics, financial markets, and industrial IoT sensors. I developed novel abstraction-based methods that could automatically learn optimal augmentation policies for different time series characteristics, significantly improving model generalization capabilities. The research involved implementing custom PyTorch modules for efficient time series transformations, designing comprehensive evaluation frameworks, and conducting extensive ablation studies. My findings were compiled into a research paper submitted to AAAI 2025, demonstrating substantial improvements in classification accuracy across multiple benchmark datasets.

Technologies Used

• Pandas
• PyTorch
• Scipy
• Weight & Biases

Supervisors

• Prof. Sylvain Kubler
• Prof. Yves Le Traon

Description

At Worldline, I worked on the development of an intelligent code analysis system that leveraged NLP techniques to understand and analyze large-scale software projects. I implemented transformer-based encoder models for sentiment analysis of code comments, commit messages, and documentation, enabling the system to detect and predict potential issues. The system featured comprehensive RAG (Retrieval-Augmented Generation) pipelines that could intelligently search through codebases, documentation, and historical data to provide contextual insights to development teams. The final system was entirely language agnostic, allowing it to be used with any programming language and framework and across projects with multiple languages.

Technologies Used

• RAG
• Langchain
• NLP
• Redis

Supervisors

• Marie-Alice Blete