AI/ML Scientist — Quantization & Numerical Robustness

Job Title

AI/ML Scientist — Quantization & Numerical Robustness

Role Summary

Research how reduced numerical precision, analog noise, and other hardware non-idealities affect modern AI models and develop quantization and robustness techniques for Arago's custom AI accelerator.

The role sits at the intersection of model research, numerical analysis, and hardware/software co-design and works closely with hardware, compiler, runtime, and inference teams to translate research into deployable solutions.

Experience Level

Mid-level (Level - Mid-Career). Specific years of experience not specified.

Responsibilities

Deliver experimental and algorithmic research to quantify and mitigate numerical and hardware-induced model degradation.

• Design and run large-scale experiments to characterize model sensitivity to analog noise, reduced precision, and hardware non-idealities.
• Develop and validate numerical and noise models representative of hardware behavior.
• Research and implement quantization approaches across training, fine-tuning, post-training, and runtime/on-the-fly methods.
• Identify model-, layer-, and operator-level precision requirements and provide actionable recommendations to hardware and software teams.
• Optimize trade-offs between model quality, numerical robustness, and inference performance.
• Collaborate with hardware, compiler, runtime, and inference engineers to integrate research into the software stack.

Requirements

Core technical skills and experience required; French is a plus.

• Deep experience with ML quantization techniques (PTQ, QAT, quantization-aware fine-tuning, mixed precision, low-bit weight/activation formats).
• Experience studying the impact of numerical precision, approximation, perturbations, or hardware noise on model accuracy and stability.
• Strong understanding of modern model architectures, such as LLMs, diffusion models, multimodal/video models, or world models.
• Ability to design rigorous, large-scale experiments and analyze accuracy/robustness trade-offs across models, layers, operators, and numerical formats.
• Good understanding of accelerator architecture, inference performance, memory/computation trade-offs, and interactions between model techniques and hardware efficiency.
• Strong Python and PyTorch skills; experience with custom operators, simulators, emerging accelerator stacks, or research prototypes is a strong plus.
• English proficiency required; French is a plus.

Education Requirements

Background in mathematics, physics, computer science, or a related quantitative field; strong foundations in numerical methods, probability, and statistics. No specific degree level or formal certification is explicitly required in the posting.

About the Company

Company: Arago

Arago is an AI and computer hardware startup (founded 2024) developing AI-driven semiconductor and photonics solutions. The company brings together engineers and researchers in photonics, electronics, software, mathematics, and machine learning to accelerate prototype-to-silicon development across hubs in France, North America, and Israel.