Principal ML Engineer

ZeroDrift, Inc.•New York, NY

5d•$200,000 - $250,000•Hybrid

About The Position

Compliance enforcement runs on specialized language models. They decide in real time whether a communication is safe to send, and they have to be accurate, fast, and reliable, because they sit in the path of live traffic. Our AI research team owns the science: model behavior, training objectives, data strategy, and the quality bar. You own the systems that make the science real. You will build the pipelines that train models reproducibly, the evaluation infrastructure that proves they work, and the serving stack that runs them in production. This is a hands-on principal individual contributor role on a small, senior team. It is a systems role, not a research role. The right candidate loves making ML industrial-grade.

Requirements

8+ years of software engineering experience, including 4+ years building infrastructure for ML or LLM systems in production
Hands-on depth with the modern LLM stack: PyTorch, distributed training, fine-tuning at scale (LoRA, SFT), and inference engines such as vLLM or TensorRT-LLM
You have built eval harnesses, regression gates, or dataset pipelines, and you understand precision, recall, and calibration well enough to build the right measurement around them
Production mindset. You have owned model serving with real latency, reliability, and cost constraints, not just notebooks
Strong fundamentals: Python, containers, CI/CD, cloud infrastructure, observability
High ownership on a small team: scope your own work, ship weekly, make pragmatic build-vs-buy calls
You enjoy being the engineering counterpart to a research partner. Tight collaboration, clear interfaces, no turf wars

Nice To Haves

Experience productionizing small or specialized language models
Experience with structured-output serving or constrained decoding in production
Prior work in a regulated or high-stakes domain such as fintech, healthcare, legal, or trust and safety
Experience deploying models into customer-controlled environments

Responsibilities

Build and own the training pipelines: data preparation, reproducible fine-tuning runs, experiment tracking, and release automation
Build the evaluation infrastructure: automated eval runs, regression gates, dashboards, and dataset versioning. Research defines what good means. You build the machinery that measures it
Own model serving in production: low-latency inference, batching, optimization, autoscaling, and cost
Ship model updates safely with versioning, canarying, rollback, and drift monitoring
Build repeatable workflows for adapting models to new domains and customer needs
Turn expert labels and reviewer feedback into clean training and evaluation data
Set the bar for ML infrastructure as the team grows