About HUD

HUD's mission is to build reliable, fair and open infrastructure for AI data. We want data to be valuable for the people who create it and trustworthy for the labs that train on it. Our team is a quickly growing group of researchers, engineers and operators building the economy that shapes what AI will become. Backed by $16M from top VCs and YC (W25), our marketplace and platform are used by startups, Fortune 500 companies and frontier labs.

About the role

We’re looking for Research Engineers to build high-quality benchmarks for evaluating frontier agents on domain-specific tasks. You’ll build benchmarks that are technically rigorous, practically useful, and credible to frontier labs.

Responsibilities

Own the design, implementation, and quality of HUD’s internal agent benchmarks

Work with subject-matter experts to define tasks and create domain-specific benchmarks that evaluate agents on realistic workflows

Build infrastructure to reliably run models and agents against benchmark tasks

Develop metrics and analyses to understand benchmark difficulty, reliability, and failure modes

Validate whether benchmark performance correlates with real-world evals, customer needs, and lab expectations

Write clear documentation and benchmark reports that make results legible and credible to technical audiences

Experience

You may be a good fit if you have

Proficiency in Python, Docker, and Linux environments

Published papers or written technical blogs on relevant topics such as public benchmarks and their limitations, model failure modes, etc. - please link in your application

Strong understanding of what a “good benchmark” means and what makes one realistic, reliable, and useful

Experience working on environments and evals

Curiosity and ability to truly understand how workflows in various domains work

Strong candidates may also

Be detail-oriented and able to spot subtle inconsistencies or edge cases in tasks

Be able to reason from first principles about task design, scoring, and failure modes

Thrive in unstructured problem spaces

Early-stage startup experience with ability to work independently in fast-paced environments

Strong communication skills for remote collaboration across time zones

We prioritize technical aptitude and learning potential over years of experience. Motivated candidates are encouraged to apply even if they don't meet all criteria.

Team & company details

Team Size: ~25 people currently, mostly full-time in-person, but some remote.

Our team: Our team brings together five International Olympiad medalists across IOI, ILO and IPhO, researchers published at NeurIPS and ICLR, and engineers and researchers from Google DeepMind, Amazon, AWS, Meta, Qwen, GLM, and Mercor. It also includes operators from BCG and Uber, the first security hire at one of the largest crypto exchanges, serial AI startup founders, former quant researchers, artists and the former #1 ranked Catan player in the world.

Company stage: We have 8 figures in funding and are scaling profitably and quickly to meet very strong demand.

Logistics

Employment: Full-time.

Location: We have offices in San Francisco or Singapore but are open to remote candidates who can work hours that 70-80% overlap with either San Francisco or Singapore time zones.

Visa Sponsorship: We provide support for relocation and visas for strong full-time candidates to the US or Singapore.

Timeline: Applications are rolling. The process is 2 technical interviews and a 2-3 day work trial.

What we offer

Competitive compensation

100% covered top-of-the-line medical, dental, and vision (US and Singapore employees)

Lunch and dinner for in-office employees

Company-wide holiday break (Christmas Eve to New Year’s Day) on top of PTO and paid holidays

401k with a 4% match

Other perks including an Equinox membership and commuter benefits (US employees) and a health/wellness stipend (Singapore employees)

Unlimited* access to tokens for ChatGPT, Claude Code, Cursor, etc. *By unlimited, we mean no one on our token usage leaderboard has ever hit a limit. So we have no idea what the limit is.