Deskripsi Pekerjaan
Shape the Future of Generative AI
At Nexus Intelligence Labs, we are building the foundational models powering the next generation of enterprise automation. We are looking for a Senior Machine Learning Engineer to join our core research team in our San Francisco headquarters. This is a high-impact role where you will own the full lifecycle of large-scale model training, from data curation and architecture design to distributed training optimization and production deployment.
Why This Role is Different
- Compute Access: Unrestricted access to 4,000+ H100 GPUs via our strategic cloud partnerships.
- Open Research Culture: Publish at top-tier venues (NeurIPS, ICML, ICLR) with dedicated legal support for IP clearance.
- Ownership: End-to-end ownership of model capabilities, not just feature engineering.
If you have a track record of scaling transformer architectures and thrive in a low-bureaucracy, high-autonomy environment, we want to talk to you.
Tanggung Jawab
- Design, implement, and optimize large-scale distributed training pipelines (PyTorch, FSDP, Megatron-LM) across thousands of GPUs.
- Architect novel model architectures (Mixture of Experts, State Space Models, Retrieval-Augmented Generation) targeting specific reasoning and coding benchmarks.
- Curate and process massive multi-modal datasets (text, code, images) using Spark/Ray, focusing on data quality, deduplication, and synthetic data generation.
- Profile and optimize kernel-level performance (CUDA, Triton, FlashAttention) to maximize MFU (Model FLOPs Utilization).
- Collaborate with Research Scientists to translate theoretical breakthroughs into robust, production-grade training runs.
- Build and maintain evaluation harnesses for alignment, safety, and capability benchmarking (MMLU, HumanEval, MT-Bench).
- Mentor junior engineers and establish best practices for reproducibility, experiment tracking (Weights & Biases/MLflow), and model checkpointing.
Kualifikasi
- Education: M.S. or Ph.D. in Computer Science, Machine Learning, or equivalent practical experience (5+ years industry experience).
- Distributed Training: Proven expertise training models >10B parameters on clusters >128 GPUs (PyTorch Distributed, DeepSpeed, Megatron, or JAX/Flax).
- Systems Programming: Strong proficiency in Python and C++/Rust; experience writing custom CUDA kernels or Triton kernels for attention/linear layers.
- Data Engineering: Hands-on experience building petabyte-scale data pipelines (Apache Spark, Ray Data, Databricks) for LLM pre-training/fine-tuning.
- Architecture Knowledge: Deep understanding of Transformer variants, normalization techniques (RMSNorm, QK-Norm), positional encodings (RoPE, ALiBi), and quantization (GPTQ, AWQ, GGUF).
- Open Source: Significant contributions to major ML open-source projects (Hugging Face, vLLM, PyTorch, JAX, Llama.cpp) or top-tier conference publications.
- Communication: Ability to articulate complex technical trade-offs to both technical and non-technical stakeholders.