Job description# **Astrophysics & Cosmology Expert** **About the Project** We're building a large-scale benchmark to test how well advanced AI systems can solve hard scientific and engineering problems. As a task designer, you'll create challenging computational problems that check whether AI can use real scientific software to do research-level work — running simulations, interpreting results, designing experiments, and uncovering hidden information from data. This isn't a typical data-labeling job. You'll design original, graduate-level problems based on real scientific workflows, test them against cutting-edge AI models, and fine-tune them until the difficulty is just right. **What You'll Do** You'll create problems that require skilled use of specialized scientific software. Some will ask the AI to compute exact answers from a fully defined setup — testing whether it can correctly carry out complex, multi-step workflows. Others will be harder: the AI must plan a series of queries or experiments to uncover information that isn't directly visible, which means thinking strategically about what to measure, how to read partial results, and how to narrow down the possibilities efficiently. Each problem goes through a testing loop against state-of-the-art AI models, and you'll refine it until it hits the target difficulty. **Domains & Tools We're Hiring For** We're especially interested in experts with deep, hands-on experience in: **Astrophysics & Cosmology** — working with astropy and related tools for cosmological calculations, angular power spectra, galaxy survey analysis, and observational data reduction pipelines. _Experience with other specialized software in this domain will also be considered._ **What Makes a Strong Candidate** You have graduate-level expertise (MS or PhD preferred) in the domain above, with real hands-on experience using these tools — not just theoretical knowledge. You've written code using these libraries to solve actual research problems, and you understand where they break, what their edge cases are, and what makes a problem genuinely hard rather than just complicated. Beyond domain expertise, the best candidates think like puzzle designers: building problems where the challenge comes from smart reasoning rather than raw computation, where several approaches seem plausible but only careful analysis reveals the right one, and where surface-level pattern matching won't get you to the answer. **Requirements** - Graduate-level training in a relevant STEM field (MS, PhD, or equivalent research experience) - Proven proficiency with at least one of the listed scientific software libraries, shown through research publications, open-source contributions, or professional work - Strong Python skills — you'll be writing problem setups, oracle functions, and solution validators - Ability to work independently and refine problem designs based on feedback - Comfortable working in a Linux/terminal environment with remote compute sandboxes - Available for at least 15–20 hours per week **Nice to Have** - Experience across multiple listed domains or tools - Familiarity with benchmark or evaluation design - Background in scientific teaching or exam/problem-set design - Experience with computational reproducibility and containerized environments **Please note:** This application includes a coding assessment as part of the evaluation process.
