Location: San Francisco, CA or Leuven, Belgium (hybrid)
Start date: January 2027
Job Description
Belfort's mission is simple: make computing on encrypted data possible. We are leading the field, working with partners like Google and NVIDIA, and Jeff Dean is one of our early backers. Our technology runs mostly on GPUs, and one of its biggest use cases is AI: running large language models on data that stays encrypted the whole time.
We are looking for an LLM Researcher to help us build and train models that work well on encrypted data. This is a research role with hands-on engineering. You will own work across the full training pipeline, from pretraining to post-training, running on large GPU clusters. You will work closely with our cryptography and engineering teams to shape models that are both accurate and efficient under encryption.
Responsibilities
Training Pipelines
Build, run and improve pretraining and post-training pipelines on large GPU clustersScale training jobs across many GPUs and nodes, and keep them fast and stableFind and fix bottlenecks in data loading, memory use, communication and throughputBuild evaluation pipelines to track model quality throughout trainingModel Research
Pretrain and post-train language models, including fine-tuning, instruction tuning and preference optimization (RLHF, DPO or similar)Adapt transformer architectures so they run efficiently on encrypted dataUse scaling laws to plan experiments and decide where compute is best spentDesign, run and analyze experiments, and turn results into clear decisionsPublish and present work where it fits our roadmapCollaboration
Work with cryptographers and engineers on the link between models and encrypted computingExplain results clearly to both technical and non-technical colleaguesQualifications
Required:
PhD in Machine Learning, Computer Science or a related field, or a Master's with equivalent research experienceHands-on experience with LLM pretraining and post-training (fine-tuning, RLHF, DPO or similar)Experience training models on large GPU clusters, including distributed training (data, tensor and pipeline parallelism)Deep knowledge of transformer architectures and scaling lawsStrong Python and PyTorch skills; you write research code others can build onStrong research track record (publications at top venues, open-source models or similar)Excellent written and verbal communication in EnglishAbility to work from San Francisco or Belgium, and available to start in January 2027Nice to have:
Experience with hardware development or hardware/software co-design (GPU kernels, FPGA or ASIC)Experience with making models smaller and faster (quantization, distillation, pruning)Knowledge of FHE or privacy-preserving machine learningApplications
Please apply through the application link for Software roles
(Note: For more general information, see the ‣)
