The work we take on.

01 /Service

Software engineering & coding tasks

Debugging, repository changes, build issues and test failures. Clear task instructions, reference solutions and checks for the expected result.

What you receive

• Task briefs and acceptance criteria
• Reference solutions and code changes
• Test cases and technical review notes

What we need to get started

Your task format, language requirements, repository access where needed, and the behaviour you want to test.

02 /Service

AI benchmarking & evaluation

Tasks that test the work a model needs to do. We define scoring criteria, review responses and document where a solution falls short.

What you receive

• Evaluation tasks and scoring rubrics
• Reviewed model outputs and error analysis
• Benchmark reports with the method and scope

What we need to get started

The model or agent workflow, evaluation goals, allowed tools and examples of the outcomes you expect.

03 /Service

Docker & terminal environments

Reproducible environments for coding and agent tasks. We check dependencies, terminal workflows and the verifiers that judge the result.

What you receive

• Docker environments and setup instructions
• Verification scripts and reference runs
• Dependency and reproducibility notes

What we need to get started

Runtime constraints, supported tools, resource limits and the format your evaluation system needs.

04 /Service

Technical review & quality assurance

Independent checks on instructions, code and evaluation results. Issues are recorded, corrections are reviewed and acceptance is documented.

What you receive

• Independent technical reviews
• Issue reports and correction records
• Checked task packages and QA notes

What we need to get started

Your acceptance criteria, task packages, review guidelines and the details needed to reproduce a result.

Data & review

Preparing data for post-training.

Clear instructions, useful demonstrations and consistent human feedback all depend on careful technical review.

We can scope supervised fine-tuning (SFT) data, human feedback (RLHF) reviews, preference comparisons and technical annotations around your guidelines. We agree on the task format, scoring rubric and reviewer calibration before delivery.

Training data and held-out evaluation sets are scoped separately. For agent work, we discuss execution environments, trajectories, verifiable rewards and checks for unintended shortcuts.