AI/ML Scientist – Quantization & Numerical Robustness

About the role

Arago is looking for an AI/ML Scientist to investigate how reduced precision, analog noise and other hardware non‑idealities impact modern AI models. You will develop quantization and robustness techniques tailored to Arago's custom AI accelerator, working at the intersection of model research, numerical analysis and hardware‑software co‑design.

Key responsibilities

• Design and run large‑scale experiments to characterize model sensitivity to analog noise, reduced precision and other accelerator non‑idealities.
• Develop and validate numerical and noise models that accurately represent hardware behavior.
• Research, implement and evaluate quantization approaches across training, fine‑tuning, post‑training and runtime.
• Identify model, layer and operator sensitivities and propose mitigation strategies.
• Collaborate with hardware engineers to align software techniques with accelerator architecture and performance constraints.

Required profile

• Strong background in mathematics, physics, computer science or a related quantitative field with solid foundations in numerical methods, probability and statistics.
• Deep experience with ML quantization techniques (PTQ, QAT, quantization‑aware fine‑tuning, mixed precision, low‑bit formats).
• Proven track record studying the impact of numerical precision, approximation or hardware noise on model accuracy and stability.
• Good understanding of modern model architectures such as LLMs, diffusion, multimodal or video models.
• Ability to design rigorous, large‑scale experiments and analyse trade‑offs across models, layers, operators and numerical formats.
• Familiarity with accelerator architecture, inference performance and memory/computation trade‑offs.
• Strong Python and PyTorch programming skills; experience with custom operators or simulators is a plus.
• Proficient English; French is a plus.

Required skills

• Python
• PyTorch
• ML quantization (PTQ, QAT, quantization‑aware fine‑tuning, mixed precision, low‑bit formats)
• Numerical methods and statistical analysis
• Experimental design for large‑scale AI research
• Accelerator architecture knowledge
• Custom operators / simulators (optional)