Senior Machine Learning Operations Engineer

About the role

Job Description About Mercury Mercury's use of machine learning in risk decisioning is growing fast in scope and in stakes. Models increasingly drive real-time decisions about fraud and financial crime, and the Machine Learning Platform (MLP) team exists to build a paved path from a trained model to a reliable production deployment, speeding up iteration, and ensuring granular production observability. MLP owns the production ML lifecycle: the systems that take a model from registry through deployment, real-time inference, observability, and retraining. Our Data Science colleagues author and t…

Requirements

• Own model deployment infrastructure: registry and versioning, CI/CD with performance, bias, and consistency checks, shadow mode, and staged rollouts
• Build model observability: availability, latency, and error monitoring, plus drift detection as a retraining trigger
• Partner with Risk Data Science to take models from a clean development-to-production handoff through to production operation under MLP ownership
• Implement experimentation capabilities such as champion/challenger and canary routing, and explainability outputs like SHAP attributions
• Feel a strong sense of product ownership and actively seek responsibility. We self-organize on small and medium projects, and we want someone excited to help shape and build a brand-new platform team
• The ideal candidate for the role has:
• 5+ years in machine learning engineering, backend software engineering, MLOps, or a closely related field
• Production ML service experience: deploying, serving, and operating models in low-latency, high-availability contexts
• Strong backend engineering fundamentals in Python, with API frameworks like FastAPI or Flask
• Experience with model deployment and lifecycle tooling: model registries, CI/CD for models, versioning, and staged rollout patterns (shadow, canary, champion/challenger)
• Experience building observability and alerting for production services: latency, errors, and ideally model-specific signals like drift
• Comfort with the data layer ML depends on: SQL, key-value/low-latency stores (Redis, DynamoDB, or equivalent), and streaming pipelines (Kafka, Kinesis, Redpanda, or equivalent)

Where this was posted

• 4 Day Week

  last seen 2h ago
  View on 4 Day Week