About the Role

Choosing between algorithmic strategies for multi-step LLM work is a measurement problem, and most teams solve it badly: comparisons run case by case, by whoever needs them that week, on whatever task is closest to hand. That doesn't scale, and it doesn't hold up to outside scrutiny - from a customer, or from a reviewer. Callosum needs one benchmarking system: reproducible, contamination-controlled, and trusted enough to be the evidence that decides which approach ships.

This role owns that system. You will build a harness that measures task success, quality, and robustness across motifs, agent topologies, and decomposition strategies, grounded in execution - real commits, real traces, sandboxed grading - rather than self-reported or model-graded scores. The results become the proof points we show customers, the evidence behind the benchmarks we co-publish, and the basis on which an approach ships or doesn't.

This is a research hire that builds. We expect the rigour of a strong evaluation paper applied to a production system, and the engineering ability to design, build, and curate it yourself rather than hand it off.

What You'll Build

• Design and build a unified system for evaluating agentic and algorithmic solutions - task success, quality, and robustness across motifs, agent topologies, and decomposition strategies, on workloads that match what customers actually run. Cost per resolved task is an outcome you track, not the object of the exercise.
• Mine real commits and traces, run sandboxed execution grading, and build task suites that reflect real agentic work: code search, code edit and repair, repository summarisation, tool use. Self-reported or model-graded success isn't enough on its own.
• Enforce controls against contamination, overfitting to benchmarks, and metric gaming, and keep baselines stable over time - any result should be re-runnable to the same number, by us or by a reviewer
• Compare algorithmic and agentic approaches honestly, not models or chips - a motif that adds steps, latency, or cost has to earn it in resolved-task quality, and the system says clearly when it doesn't
• Lead external benchmark co-publications, held to a standard that survives peer and customer review
• Feed results directly into which approach ships, into the proof points behind customer engagements, and review quality claims across the company before they go out

What You'll Bring

• PhD in computer science, machine learning, or a related field, or an equivalent research track record
• Authorship or co-authorship of a benchmark or evaluation paper at a recognised venue - NeurIPS Datasets and Benchmarks, ICML, ICLR, ACL - ideally on agentic or LLM evaluation, or a comparably rigorous evaluation contribution
• A working understanding of how LLM and agent evaluation goes wrong: contamination, overfitting to benchmarks, weak baselines, underpowered comparisons, irreproducible results
• The engineering ability to design, build, and curate these systems decisively - strong Python, and comfort with sandboxed and distributed execution and CI
• Hands-on experience building or rigorously evaluating agentic or multi-step LLM systems

What Sets You Apart

• Published agentic or tool-use benchmarks that use execution-based grading
• Experience running sandboxed execution grading at scale
• Open-source evaluation or harness tooling
• Familiarity with code-agent workloads such as search, edit, and repair

What We Offer

• Competitive Salary, determined by skills and experience
• Equity & Ownership
• Private healthcare
• We offer Visa sponsorship and relocation benefits to hire the best in the world
• We work in person at our London office. You'll have the tools, space and setup to do your best work, and if you have specific needs, just tell us
• We're committed to building an inclusive workplace where everyone feels welcome, and believe in equal opportunities for all.