Principal Machine Learning Engineer

Roche•New York, NY

4d•$192,500 - $357,500•Onsite

About The Position

Advances in AI, data, and computational sciences are transforming drug discovery and development. Roche’s Research and Early Development organizations at Genentech (gRED) and Pharma (pRED) have demonstrated how these technologies accelerate R&D, leveraging data and novel computational models to drive impact. Seamless data sharing and access to models across gRED and pRED are essential to maximizing these opportunities. The new Computational Sciences Center of Excellence (CoE) is a strategic, unified group whose goal is to harness the transformative power of data and Artificial Intelligence (AI) to assist our scientists in both pRED and gRED to deliver more innovative and transformative medicines for patients worldwide. At Roche's AI for Drug Discovery (AIDD) group (Prescient Design), we are revolutionizing drug discovery with cutting-edge machine learning. We are seeking a Principal Machine Learning Engineer to join our Foundation Models team. In this role, you will drive the engineering, scaling, and operationalization of our internal reasoning Large Language Models (LLMs) and agentic systems, enabling them to succeed at complex biomolecular design and autonomous scientific workflows. You will work at the intersection of engineering and research, spanning the full stack: from the agent orchestration logic that makes these systems scientifically useful, to the distributed infrastructure and MLOps/AgentOps that make them robust at scale.

Requirements

BS, MS, or PhD in Computer Science, Machine Learning, Engineering, or a related quantitative field.
Demonstrated track record of technical leadership with increasing levels of experience based on degree: PhD with 5+ years, MS with 8+ years, or BS with 10+ years of industry experience building, shipping, and owning large-scale ML systems and infrastructure end-to-end.
Exceptional Python programming skills and rigorous software engineering fundamentals (Git, automated testing, CI/CD, documentation, architecture design).
Extensive hands-on experience with modern deep learning frameworks (PyTorch, JAX) and deploying ML infrastructure on AWS or HPC environments, including distributed training tools.
Practical experience designing agent orchestration frameworks (e.g., LangGraph, MCP-based tool integration), managing persistent agent memory, and building self-improving loops.
Strong passion for applying frontier AI and agentic science to AI for Drug Discovery (AI4DD), biology, and chemistry.

Nice To Haves

Deep expertise in LLM serving, test-time compute, sampling/search strategies, model routing, batching, caching, and latency/cost/quality tradeoffs.
Experience working with molecular modalities (e.g., protein sequences, chemical graphs, and structured molecular data).
A public portfolio of significant technical contributions to open-source ML, systems, or MLOps libraries.

Responsibilities

Architect and deploy autonomous agents that utilize tools, retrieve scientific evidence, and execute multi-step reasoning across drug discovery workflows.
Design and implement advanced agent memory architectures and context management for long-horizon scientific tasks.
Build reliable interfaces between agents and genomic, chemical, and clinical data sources.
Design, build, and optimize large-scale distributed training and inference systems for foundation models.
Own the production Python/PyTorch codebases that turn fast-moving research into enterprise-grade software.
Establish best practices for the full lifecycle: experiment tracking, system observability and monitoring, rigorous evaluation harnesses (checking agent output against scientific ground truth), CI/CD, and infrastructure.
Define the long-term engineering roadmap for AI4DD’s agentic and foundation models.
Serve as a technical authority on ML infra for Genentech leadership, architect cross-functional platforms, and elevate the engineering bar across gRED.
Partner closely with ML Scientists and domain experts to translate open-ended scientific problems and complex reasoning objectives into scoped, shippable, and highly efficient systems.