Senior AI Researcher, Evaluation & Agent Performance

Cobalt; New York (NY); On-site; USD 160,000 - 230,000; Full time

Job summary

Cobalt is seeking an experienced Senior AI/ML researcher to lead evaluation and benchmarking of its agent ecosystem, including code Q&A and MCP tooling. You will design eval harnesses, ground truth from production data, and measure recall, precision, and efficiency across models and prompts.

The role requires 7+ years in ML/NLP research, deep knowledge of agent architectures, and production-grade Python tooling.

Qualifications

• 7+ years in ML/NLP research or applied AI with evaluating LLMs or agents.
• Experience designing benchmarks: task sampling, contamination control, scoring methods, and power analysis.
• Deep knowledge of agent architectures: tool use, retrieval, multi-step planning, MCP or similar protocols.
• Expert Python, and production-quality eval infrastructure you have shipped.
• Experience leading research projects end to end, from design to publication or product.

Responsibilities

• Own the evaluation function for all Cobalt agents: code Q&A, change-impact analysis, incident response, and MCP tool use by coding agents.
• Design and build eval harnesses and benchmarks, with ground truth taken from merged PRs, production traces, and code.
• Define and track agent performance metrics: recall and precision, cost per correct answer, tool-call efficiency, latency, and failure modes.
• Run controlled comparisons across models, prompts, retrieval strategies, and baselines, with statistical rigor.
• Diagnose agent failures in retrieval, reasoning, and tool design, and drive the fixes with engineering.
• Build regression suites in CI that score every model, prompt, and connector change before release.
• Lead external research, including ownership of research papers.
• Set the research agenda and mentor future hires.

Skills

7+ years in ML/NLP research or applied

Benchmarks design

Agent architectures knowledge

Expert Python

Experience leading research projects

Tools

Python

CI/CD pipelines

Experiment tracking

Job description

We build the context layer for engineering organizations. We connect code, runtime telemetry, data lineage, and operational systems into one live model of how software runs, then give AI agents that model through chat and MCP.

What you'll do

• Own the evaluation function for all Cobalt agents: code Q&A, change-impact analysis, incident response, and MCP tool use by coding agents.
• Design and build eval harnesses and benchmarks, with ground truth taken from merged PRs, production traces, and code.
• Define and track agent performance metrics: recall and precision, cost per correct answer, tool-call efficiency, latency, and failure modes.
• Run controlled comparisons across models, prompts, retrieval strategies, and baselines, with statistical rigor.
• Diagnose agent failures in retrieval, reasoning, and tool design, and drive the fixes with engineering.
• Build regression suites in CI that score every model, prompt, and connector change before release.
• Lead external research, including ownership of research papers.
• Set the research agenda and mentor future hires.
  

Requirements

• 7+ years in ML/NLP research or applied AI, including evaluating LLMs or agents.
• A track record of designing benchmarks: task sampling, contamination control, scoring methods, and power analysis.
• Deep knowledge of agent architectures: tool use, retrieval, multi-step planning, MCP or similar protocols.
• Expert Python, and production-quality eval infrastructure you have shipped.
• Experience leading research projects end to end, from design to publication or product.

Nice to have

• Publications at NeurIPS, ICLR, ACL, ICSE or FSE on LLM evaluation, code intelligence or agents.
• Experience with repo-level code benchmarks (SWE-bench-style).
• Experience with LLM-as-judge methods and their known failure modes.
• A background in distributed systems, observability data, or large codebases.
• Enterprise ground truth: thousands of repos, live traces and production incidents.
• Your evals decide what ships.
• A founding role in the research function.