Deskripsi Pekerjaan
Nexus Intelligence is building the foundational infrastructure for the next generation of autonomous AI agents. Backed by top-tier VCs (Sequoia, a16z), we are a team of 45 engineers and researchers pushing the boundaries of LLM reasoning, tool use, and long-context memory architectures.
We are looking for a Senior Machine Learning Engineer to own the end-to-end lifecycle of our core embedding and retrieval models. You will not just fine-tune existing models; you will design novel training pipelines, optimize inference for latency/cost at massive scale, and directly shape the 'brain' of our agent platform.
This is a high-autonomy, high-impact role for someone who wants to move fast, publish great work, and see their models power millions of autonomous workflows daily.
Tanggung Jawab
- Design, train, and deploy state-of-the-art embedding and reranking models optimized for RAG and agentic tool retrieval.
- Architect scalable data curation pipelines (synthetic generation, human feedback loops, hard negative mining) using tools like Ray, Spark, or custom Kubernetes operators.
- Optimize model inference latency and throughput (TensorRT, vLLM, ONNX Runtime, Triton) for real-time agent loops.
- Conduct rigorous offline evaluation (benchmarks, A/B testing frameworks) and online experimentation to measure business impact.
- Collaborate with Product and Platform teams to define model requirements for new agent capabilities (e.g., code execution, browser automation, long-horizon planning).
- Stay current with literature (ICLR, NeurIPS, ICML) and implement relevant techniques (e.g., Matryoshka Representations, Colbert, speculative decoding) into production.
- Mentor junior engineers and establish best practices for MLOps, model monitoring, and reproducible research.
Kualifikasi
- 5+ years of industry experience training and deploying deep learning models in production environments.
- Expert proficiency in PyTorch/Jax and the Hugging Face ecosystem (Transformers, Accelerate, PEFT, TRL).
- Proven track record of optimizing LLMs or Embedding models for latency/cost (Quantization, Distillation, Pruning, FlashAttention).
- Deep understanding of Retrieval-Augmented Generation (RAG) architectures, vector databases (Milvus, Weaviate, Pinecone), and evaluation metrics (nDCG, Recall@K, MRR).
- Strong software engineering fundamentals: Python, Git, CI/CD, Docker, Kubernetes, and cloud infrastructure (AWS/GCP).
- Experience with large-scale distributed training (FSDP, DeepSpeed, Megatron-LM) on multi-node GPU clusters.
- Publications at top-tier venues (NeurIPS, ICML, ICLR, ACL, EMNLP) or significant open-source contributions to ML libraries are a major plus.