Iheb Gafsi

R&D Engineer | ML Engineer

Developing rigorous tools from physics, mathematics, and computer science to understand how large neural networks learn, compute, scale, reason, and imagine.

About

I instrument and develop rigorous tools from physics, mathematics, computer science, and theoretical neuroscience to understand how large neural networks learn, compute, scale, reason, and imagine. I concomitantly use machine learning tools to understand physical systems and their behavior, like particle physics and quantum mechanics.

Machine Learning
Mathematics
Reinforcement Learning
Statistical Physics
Quantum Computing
Scientific Computing
Large Language Models
Neural Networks

Experience

Machine Learning Engineer

NanoTensor
Oct 2025 - Present

Employing and developing tools from mathematics and physics in computational Physics, Large Neural Networks, and Pattern Recognition for scale.

Technical Course Writer

DataCamp
Jan 2025 - Present

Creating educational materials on Artificial Intelligence and Machine Learning, focused on breaking down complex technical concepts for millions of learners worldwide.

Deep Learning Instructor

NVIDIA
May 2024 - Nov 2025

Teaching courses on Efficient Large Language Models Customization, Computer Vision for Industrial Inspection, Fundamentals of Machine Learning, and Building Transformer Based NLP systems.

Scientific Machine Learning Intern

CAMMP
Jul 2025 - Oct 2025

Research in machine learning applications for materials science and manufacturing processes.

R&D Machine Learning Engineer

Callem AI
Apr 2024 - Nov 2024

Research and development in multi-agent systems, AI agents, and distributed systems.

Research Projects

Gaussian B-Splined Kolmogorov Arnold Physics Informed Neural Networks

Jul 2025

Reinforced the power of Kolmogorov-Arnold approximators for interpretable machine learning by utilizing Gaussian B-splines. Demonstrated universality of KANs as a generalization converging to mathematical approximators like Fourier Series and Taylor Series.

#machine_learning #scientific_computing #navier_stokes

Optimizing Restricted Boltzmann Machines via Quantum Annealing

Jan 2025 - May 2025

Proposed a quantum restricted Boltzmann machine (QBM) that creates accurate quantum deep memory representations via quantum annealing and quantum Boltzmann sampling. Addressed computational complexity limitations in high-dimensional Hilbert spaces.

#quantum_computing #statistical_physics #machine_learning

Kolmogorov Arnold Neural Networks Framework

Jan 2025

Python implementation of the Kolmogorov-Arnold Neural Networks Framework based on the Kolmogorov-Arnold Representation Theorem for regression and classification tasks.

#neural_networks #framework

Publications

RAYGo: Retrieve As You Go for Large-Scale Asynchronous Action Planning

Zenodo · Sep 2024

Introduced a novel approach combining Chain-of-Actions (CoA) and Retrieve-As-You-Go (RAYGo) to enhance LLM agent planning. Enabled Mistral-7B-v0.2 to outperform GPT-4 by over 29% in multi-tool calling, and elevated GPT-4's performance from 56% to 98% on action calling.

Linear-Time Sequence Modeling with Selective State Space Mamba

Zenodo · Jan 2023

Presented Mamba, a new architecture leveraging structured state space models that scales linearly with sequence length, achieving 5× better throughput than Transformers while maintaining state-of-the-art performance across languages, audio, and genomics.

Transformers

Zenodo · Aug 2023

Comprehensive overview of the Transformer architecture, its self-attention mechanism, and its role in revolutionizing natural language processing and modern machine learning applications.

Education

INSAT - Institut National des Sciences Appliquées et de Technologie

Engineering, Applied Sciences and Technology

Aug 2022 - Jun 2027

ACM, IEEE, GDSC, INSAT Press

Get in Touch

Interested in collaboration or have questions about my research? Feel free to reach out.