●Jebra Research
Payroll rules are written in prose. Payroll systems need logic.We're fine-tuning a language model to convert natural-language pay agreements into structured, auditable payroll rules — the step before calculation.
The distinction
Jebra.io calculates payroll.
Jebra.ai extracts the rules.
Jebra.io
Our payroll calculator for shift-based businesses. It runs the numbers once the logic is known. // the calculation
Jebra.ai
The step before calculation. It turns pay agreements into that logic. // the extraction
The problem
In shift-based work, the rules aren't structured yet.
Pay rules don't live in a database. They sit across pay agreements, contracts, spreadsheets, scheduling systems, manager notes and payroll history.
A human payroll manager can interpret that context. A payroll system cannot use it — until it becomes structured logic.
The model
Trained for extraction, not conversation.
A language model fine-tuned on a single task: reading a pay agreement and producing a rule a payroll engine can execute.
01 Extract
Rates, units, premiums, caps and exceptions — pulled out of prose and named as fields.
02 Map
Pay terms mapped onto a structured payroll schema, so the same rule means the same thing every time.
03 Detect
Ambiguity flagged before a rule is used — not after it has already been paid.
Research
We publish the work as we do it.
Model performance on real and synthetic pay agreements, where off-the-shelf models fail, and how fine-tuning improves accuracy — including the results that don't go our way.
Write for the machine
How abbreviation, syntax, redundancy and filler change what a model can pull out of a pay agreement.
Read on dev.to ↗When payroll gets complicated, the machine gets harder to please
We made the agreements as complex as real shift work. The clean result did not survive.
Read on dev.to ↗Dicts and docs: the value of grammar and documentation for LLM-based automation
Structured dictionaries and documentation change what a model can reliably extract from a pay agreement.
Read on dev.to ↗