When we released Holo4 on 28 September, we published every trajectory behind its public benchmark scores at trajectories.hcompany.ai. For OSWorld that's 369 desktop tasks, each run 3 times, with the steps, tokens and time of every attempt. This role builds and runs the evaluation framework that produces runs like these.
H builds computer-use agents and the models behind them. Developers use them through a managed API, and our forward deployed engineers take them into enterprise workflows.
What this team ownsThe evaluation framework: orchestration, runtimes and observability. Researchers and forward deployed engineers bring the benchmarks, across web apps, desktop applications and the command line. Your job is to make the framework that runs them reliable, fast and cheap, and to make adding a new one quick. Research uses the results to choose checkpoints and decide whether a model ships. Product and the forward deployed engineers use them to measure agents on customer workflows. It carries roughly 50 benchmarks now. That number should be between 100 and 200 soon, and the framework has to keep up.
What you'd be doingIntegration support for researchers and forward deployed engineers bringing in a benchmark, with a shorter path each time.
Setting the standard for how a benchmark enters the framework, and building the checks that enforce it.
Scheduling and observability, so cluster capacity isn't left idle while evaluation jobs queue.
Reproducible results across trials, so a release decision rests on numbers that hold.
Whatever stack a benchmark calls for. One week that's cluster tuning; the next it's a browser extension or desktop environments.
Time with customers, from single developers to large companies, to find out what they want measured, then automating it so the results flow back into our harnesses and models.
By 3 months you'll have helped researchers or forward deployed engineers integrate 5 benchmarks, and started fixing what slows the framework down. By 6 months one part of it is yours, for example scaling the runs, observability, or a group of related benchmarks, and a release will have gone out on your numbers. By 12 months you'll know the design and trade-offs of the whole evaluation system, and be the person the rest of H asks about evaluations.
Who you'd work withCeiran Chapman, our VP Engineering, is hiring for this role. You'd join the evaluation team. The people relying on your work day to day are H's researchers and forward deployed engineers.
What we think it takesLikely a good fit if you
Have spent 5+ years in backend development, with production Python at the core, and use coding agents to go faster without letting quality drop.
Have built test, QA or evaluation tooling that other teams depended on, and care whether a number is right.
Have operated distributed systems on Kubernetes in a public cloud. AWS experience helps most.
Have built and shipped systems end to end, including APIs (REST or GraphQL) and integrations with outside services.
Know relational and non-relational databases, and message queues such as SQS, RabbitMQ or Kafka.
Instrument what you build, with metrics, tracing and monitoring from the start.
Stronger still if you have
Measured LLM quality before, or built agents yourself.
Packaged and run workloads in Docker and on virtual machines.
Used Temporal, Dask, FastAPI, PostgreSQL, Grafana or Datadog.
Automated web or desktop software with Playwright, Selenium or a browser extension you wrote.
Set standards other engineers follow, through code review, design review or mentoring.
You do not need a background in machine learning. We'll work that out with you. If you match most of this but not all of it, apply anyway.
How we hireA 30 minute call with our Talent team, a 60 minute technical challenge, a 60 minute system design interview, and a 30 minute final conversation with Ceiran. About 3.5 hours in total.
PracticalitiesLondon posting: Hybrid in London. That means 3 office days a week and a Paris trip about once every 4 to 6 weeks. There is a Paris posting for the same role. We offer a competitive package.
