About the Role
Please submit your CV in English and indicate your level of English proficiency.
Mindrift (https://himalayas.app/companies/mindrift) connects specialists with project-based AI opportunities for leading tech companies, focused on testing, evaluating, and improving AI systems. Participation is project-based, not permanent employment.About the Role
You'll design coding tasks that challenge frontier AI coding agents. Each task is a self-contained Docker environment with a broken piece of software; an AI agent attempts the fix; automated tests verify the outcome. Your deliverable is the full task package: broken code, tests, instructions, and a reference solution proving the task is solvable.
Responsibilities:
Invent a realistic developer scenario — a real bug, a broken ETL, a missing feature — not a toy problem.
Build a reproducible Docker environment with pinned dependencies.
Write a pytest that verifies outcomes, not specific commands — deterministic, non-flaky, and does not leak the fix.
Write an instruction.md that reads like a Jira ticket a developer would receive.
Write a reference solve.sh proving the task is solvable.
Calibrate difficulty so current state-of-the-art agents solve the task 20–60% of the time.
Iterate based on feedback from expert QA reviewers.
Later: review other authors' tasks as a QA reviewer.
Not in scope
Data labeling, prompt engineering.
Production code to ship — you design problems and verification for AI agents.
Leetcode puzzles — scenarios must look like real developer work.
Not every candidate task ships — quality over quantity.
Requirements
3+ years of production software development in one backend stack — Python, Go, Node.js, Java, or Rust. Depth in one stack beats breadth.
Python + pytest fluency — required regardless of primary stack. The task harness is pytest-based even when the broken app is in another language. Fixtures, parametrize, monkeypatch, timeouts, conftest.py.
Docker authoring — reproducible Dockerfiles, pinned dependencies, multi-stage builds when needed, non-root user.
Linux & Bash — comfort debugging inside containers (strace, lsof, journalctl); shell beyond set -euo pipefail.
AI coding agent experience — Claude Code, Cursor, Roo Code, or similar, on non-trivial work. You can cite a specific time the AI was confidently wrong and how you caught it.
English — B2+ written.
Not a fit
Data Science, ML, or Computer Vision engineers without backend-engineering output.
Manual QA testers without automation or test authoring.
Frontend-only, low-code / no-code, IT Support, or Business Analysts.
Engineers who have never written pytest from scratch.
Junior, intern, or assistant as the most recent role.
Preferred qualifications
Domain depth in Security, System Administration (nginx / systemd / cron), Scientific Computing (NumPy / PyTorch / SciPy), DevOps, or Git internals.
Modern Python tooling (uv, poetry, pyproject.toml).
Coverage tooling (pytest-cov, coverage.py, gcov, llvm-cov, kcov).
Fuzzing or property-based testing (Hypothesis).
Prior contribution to agent-evaluation benchmarks or related frameworks.
Process
Apply → Pass qualification (90-minute sample-task screen + short behavioral interview) → Join a project → Complete tasks → Get paid.Time commitment
Onboarding: ~10 hours per first task.
Steady state: ~5 hours per task, 2–4 parallel tasks per author.
Realistic weekly load: 8–20 hours. Higher volume available for top performers.
You choose when and how to contribute; tasks must be submitted by the deadline and meet acceptance criteria.
Compensation:
Paid contributions, rates up to $35/hour*.
Task-based compensation equivalent to hourly rate, depending on performance and volume.
Some projects include incentive payments.
*Rates vary based on expertise, skills assessment, location, project needs, and other factors. Higher rates may be provided to highly specialized experts. Lower rates may apply during onboarding or non-core project phases. Payment details are shared per project.Apply
Submit your CV via the Mindrift (https://himalayas.app/companies/mindrift) platform. Indicate your English level, note this role (Software Engineering Evaluation Specialist — Terminal Bench), and include a GitHub profile link if available.
Originally posted on Himalayas (https://himalayas.app)
A name is a term used for identification by an external observer. They can identify a class or category of things, or a single thing, either uniquely, or within a given context. The entity identified by a name is called its referent. A personal name identifies, not necessarily uniquely, a specific individual human. The name of a specific entity is sometimes called a proper name and is, when consisting of only one word, a proper noun. Other nouns are sometimes called "common names" or (obsolete) "general names". A name can be given to a person, place, or thing; for example, parents can give their child a name or a scientist can give an element a name.