Machine Learning Engineer

A high-growth AI-native startup is hiring an ML Engineer in New York to take models from notebook to production and keep them performant, reliable and affordable at scale. This role sits at the intersection of ML and engineering: strong Python, strong PyTorch, and the engineering maturity to turn models into real product surfaces. You’ll own training, serving, evaluation and the plumbing underneath.

Role

ML Engineer Salary: $180,000 - $300,000 plus equity Location: New York, 5 days per week in office What you’ll be doing: Train, fine-tune and serve models in production using Python and PyTorch Build and maintain the inference stack: model serving (vLLM, Triton, custom), batching, caching, cost control Stand up training infrastructure using Ray, Modal, SageMaker or similar Work with vector databases (pgvector, Pinecone, Weaviate) for retrieval-heavy workloads Build evaluation pipelines so model quality is measurable rather than anecdotal Partner with product and AI engineers to turn research or OSS models into shipped features Own GPU utilisation, latency and inference cost as first-class concerns Use AI tooling (Claude, Cursor and similar) across every part of your workflow What you’ll need: Strong Python and PyTorch experience with real production ML systems behind you Experience with Hugging Face Transformers, modern training workflows and distributed training Hands-on with model serving stacks (vLLM, TensorRT, Triton or equivalent) and GPU-level optimisation Comfort with experiment tracking (MLflow, Weights & Biases) and reproducibility hygiene Cloud experience (AWS, GCP or Azure) and modern infra tooling Fluent in modern AI engineering tools and able to use them to raise output Happy working 5 days a week in-person with a small, high-intensity team Nice to have: Experience fine-tuning or post-training LLMs (SFT, DPO, RLHF) Low-latency inference work or custom CUDA kernels Prior work securing or hardening ML systems in production This role would suit an ML engineer who treats the inference bill, the latency graph and the eval harness as part of their job, not someone else’s.