AI Research Engineer, Safety Pretraining for Open Frontier Models, EPFL (x3)

Lausanne, Switzerland

• In this role, you'll build the safety-pretraining stack for Apertus, scaling alignment methods into production LLM pipelines.
• Develop distributed training infrastructure using Megatron-LM to optimise throughput across thousands of GPUs.
• Design synthetic-data pipelines with teacher-LLM orchestration, quality control, and adaptive curriculum systems.
• Build evaluation harnesses with inference optimisation, LLM-as-a-judge infrastructure, and automated jailbreak testing.
• Integrate safety alignment throughout pretraining to weave values into model training from the start.

The Swiss AI Initiative is an open science/open source effort for AI foundation models.