Marouane Talaa
AI Research Scientist & Machine Learning Engineer
Paris, FR.About
Highly analytical and innovative AI Research Scientist with a strong academic foundation in Mathematics and Computer Science, specializing in Large Language Models (LLMs), data attribution, and real-time AI system deployment. Proven ability to develop and optimize complex machine learning solutions, from cutting-edge research to scalable production APIs, driving significant advancements in information extraction, speech-to-text, and predictive analytics. Eager to leverage deep expertise in AI/ML, statistical modeling, and cloud technologies to solve complex challenges and contribute to pioneering projects.
Work
→
Summary
Leading advanced research in AI, focusing on information extraction and large language model interpretability to develop innovative solutions and drive data attribution capabilities.
Highlights
Developed novel methods for information extraction from visually rich documents, enhancing data processing efficiency and accuracy.
Conducted mechanistic probes of Large Language Models (LLMs) like PleIAs, GPT2, and Bloom using scikit-learn, providing critical insights into residual stream interpretation across layers.
Optimized state-of-the-art influence functions on LLMs, enabling large-scale experimentation and performance benchmarking for various formulations (sequence-wise and token-wise).
Executed Distributed Data Parallel (DDP) experiments in PyTorch on large datasets (> 1 billion tokens), significantly accelerating model training and scalability.
Implemented a caching and retrieval framework to streamline the data attribution task, improving overall research workflow efficiency.
→
Summary
Consulted on AI model deployment and API development, delivering robust solutions for interactive agent systems with real-time text and voice capabilities.
Highlights
Developed a high-performance API for a role-playing AI Agent, integrating text responses with session history and voice responses via XTTS voice cloning.
Engineered and deployed the agent API using FastAPI on Runpod GPU instances, ensuring scalable and efficient real-time processing capabilities.
→
Summary
Developed and deployed AI-driven applications, focusing on speech-to-text and leveraging cloud-based platforms for data management and model optimization.
Highlights
Developed an Android Application Proof-of-Concept (POC) for speech-to-text, integrating an embedded on-device model for enhanced offline capability and responsiveness.
Utilized Google Cloud Platform (GCP), including Vertex AI, for model experimentation, fine-tuning, and robust cloud data management, optimizing development workflows and data pipelines.
Education
Languages
French
English
Skills
Programming Languages
Python, C++, R.
Machine Learning & AI
Machine Learning, Deep Learning, Natural Language Processing (NLP), Large Language Models (LLMs), Transformers, Data Attribution, Influence Functions, Distributed Data Parallel (DDP), Model Deployment, Speech-to-Text, Bayesian Statistics, Maximum Entropy, Regression, Kernel Density Estimation, PLS, PCR, Sparsity, Dual Norm, Regularisation, Maximisation-Minimisation, Vector Quantization, k-means, Hebbian Learning, Time Series Analysis, LSTM, Ensemble Methods.
Libraries & Frameworks
scikit-learn, PyTorch, FastAPI, OpenCV, xtts.
Cloud Platforms & Tools
Google Cloud Platform (GCP), Vertex AI, Runpod, Data Pipelines, Benchmarking, Statistical Analysis, Algorithm Design.