SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Salary: USD 60 - 90 / hourly
Huzzle Labs partners with frontier AI labs to evaluate and improve frontier models through deep human expertise. This role focuses on creating realistic accounting tasks that test how well advanced AI systems handle accounting work, and developing scoring rubrics to assess model performance consistently.
You will write realistic accounting prompts—the kind a client, controller, or business owner would send—covering month-end close, bank and credit card reconciliations, AR/AP ageing, accruals and adjusting entries, sales tax, job costing, and cleanup of incomplete or inconsistent books. You will review deal rooms containing hundreds of files, select relevant documents, and develop prompts and rubrics for tasks where frontier models struggle.
Each request should call for a usable deliverable, typically a spreadsheet, journal entry schedule, or short memo. You will test each prompt against a frontier model, reviewing the output with the same care you would apply when signing off accounting work. Where the model succeeds, you will increase difficulty until a genuine gap appears.
You will write objective, checkable scoring rubrics that another reviewer can apply without repeating your work—for example, "the reconciled ending balance equals $X" and "the $X deposit is recorded as a customer deposit liability, not revenue." Your judgment helps define the standards used to measure model performance. Training is provided, and an NDA must be signed before engagement begins.
Requirements:
- At least 5 years in accounting, bookkeeping, or finance operations
- Prior AI training or data-annotation experience is required
- Hands-on knowledge of month-end close, reconciliations, AR/AP, accruals and adjusting entries, sales tax, and job costing
- Excellent written English and precision with numbers
- Careful and detail-oriented; accurate figures and objective, checkable criteria
- Available 20 to 40 hours per week for an initial one-week trial, with potential extension
