Research Engineer, Large Language Models
Must-have:PythonNode.jsKubernetesCloudDataAI
Key Responsibilities – What You'll Do
- Design and run large-scale pre-training and fine-tuning experiments, from small ablations to full training runs.
- Study scaling laws to choose model size, data mix, and hyperparameters under a fixed compute budget.
- Improve training efficiency through parallelism strategies (data, tensor, pipeline, FSDP), mixed precision, and custom CUDA or Triton kernels.
- Diagnose and fix training instabilities such as loss spikes, divergence, and hardware failures.
- Build and improve data pipelines, including filtering, deduplication, quality scoring, and dataset mixing.
- Implement ideas from recent research papers and test whether they hold at scale.
- Build evaluations that measure real model capabilities and track results rigorously.
- Work closely with the alignment, infrastructure, and product teams, and share findings through clear write-ups.
Required Qualifications – What We're Looking For
- 3+ years of hands-on experience training deep learning models, or a PhD in a related field (computer science, mathematics, or physics).
- Strong Python skills and deep expertise in PyTorch or JAX.
- Practical experience with Transformer architectures, including attention, tokenization, embeddings, and positional encodings.
- Experience with distributed training on multi-node GPU clusters (FSDP, DeepSpeed, Megatron-LM, or similar).
- Solid foundations in linear algebra, probability, statistics, and optimization.
- Ability to read a research paper, implement it from scratch, and judge whether it works.
- Rigorous experimental habits, including clean baselines, ablations, reproducible results, and experiment tracking.
- Clear written and verbal communication skills, and comfort working in a fast-moving, ambiguous environment.
Preferred Qualifications – Nice to Have
- Publications at top venues (NeurIPS, ICML, ICLR, ACL) or well-known open-source contributions.
- Experience pre-training an LLM from scratch, at any scale.
- Post-training experience, including supervised fine-tuning, RLHF, DPO, or other preference-optimization methods.
- GPU performance work involving CUDA, Triton, FlashAttention-style kernels, and profiling.
- Experience with efficient inference, including quantization, distillation, and optimized model serving.
- Cluster and infrastructure experience with Slurm, Kubernetes, or cloud GPU platforms.
- Large-scale data engineering experience (Spark, Ray, dataset curation for multilingual or domain-specific corpora).