Develop production ML systems that score and refine outputs where correctness is nuanced and labels are imperfect.
Architect evaluation frameworks for tasks where ground truth is partial, delayed, or disputed.
Create feedback loops that translate reviews, disagreements, corrections, and adjudication into measurable model and system improvements.
Own end-to-end production ML behavior, including precision and recall tradeoffs, regression detection, drift, latency, cost, and explainability.
Enhance model quality using appropriate approaches such as prompting, fine-tuning, retrieval, active learning, heuristics, and error analysis.
Collaborate with backend engineers to integrate inference into durable, long-running workflows while preserving debuggability and human oversight.

Proven track record shipping ML systems that improved a real product, workflow, or business metric.
Strong instincts for model quality, evaluation design, error analysis, and production failure modes.
Comfort operating in ambiguous problem spaces with imperfect labels and evolving correctness.
Judicious decision making about when to apply prompting, fine-tuning, retrieval, human review, or simpler constraints.
Solid engineering fundamentals across the full ML stack, not limited to modeling.
Familiarity with LLM applications, model-assisted workflows, evaluation frameworks, or human-in-the-loop ML is a strong plus.
Preference for simple, inspectable ML systems that improve quickly and fail in understandable ways, rather than overly complex architectures.
Comfort deploying models only when there is a clear evaluation story.
Ability to navigate ambiguity without paralysis and make reasonable bets with incomplete information.
Focus on real-world impact and output of the system, not solely benchmark metrics.

Move quickly within a young, high-ownership codebase where decisions have long-term architectural impact
Work across models, data, backend systems, and product surfaces, with frequent context switching
Debug production ML failures in live, long-running workflows where silent errors matter
Collaborate closely with backend engineers on a stack that includes Python, Temporal, Postgres, AWS, and LiteLLM
Balance automation confidence with human review, recognizing when to defer versus ship

The architecture is not fixed; early engineers define how quality is measured, how models and humans interact, where automation is trusted, and how the system compounds over time
The feedback loop is short, with visible impact on what customers receive when model behavior changes ship
You operate on a strategically central product area at Mercor during a moment when frontier AI challenges lack robust solutions

Machine Learning Engineer, Frontier Data Products

Job Description