Manual sorting hides uncertainty in the pile.
Payslips can arrive as PDFs or images with inconsistent filenames. Sorting them manually requires opening each file, identifying dates and deciding whether a document is missing or repeated.
The prototype tackles this bounded organisation task. Its value proposition is a hypothesis to test, rather than a demonstrated claim about time saved or extraction accuracy.
Extract a few fields, then organise the batch.
The core flow accepts multiple files, sends each document for AI-assisted classification and extraction, then uses the returned payment date and identity fields to organise the results.
The interface exposes per-file processing states and an analysis report before offering a combined PDF. Keeping the experiment close to one end-to-end task makes its assumptions easier to inspect.
Connect extraction, sorting and PDF output.
The React application accepts PDFs and common image formats. A Gemini integration requests structured fields for payment date, employee, employer, pay period and amount, alongside a payslip classification.
The application sorts successful results by date, reports repeated dates and apparent missing months, then uses pdf-lib to combine PDFs and supported images. Per-file removal can regenerate the report and pack for the remaining set.
A simple rule can be too confident.
The prototype treats matching dates as duplicates and retains one document per date in the combined pack. That is a useful implementation shortcut to examine, but matching dates alone do not establish that two payslips are duplicates.
Its monthly-gap heuristic also assumes a cadence that may not match weekly or fortnightly payroll. Extraction uncertainty and legitimate multiple employers make human review and stronger identity checks important next design considerations.
Make the uncertain cases visible.
The result demonstrates an AI-assisted organisation loop, not validated payroll intelligence or a production document service. No benchmark for extraction accuracy is recorded in the reviewed sources.
Proposed next steps are editable extracted fields, explicit inclusion decisions, payroll-aware gap analysis and server-side handling of AI credentials. Representative document testing should precede broader use. The broader SkillAssist project addresses related review and workspace concerns, but a direct code migration between the projects is not established.
What exists. What comes next.
A prototype that extracts payslip fields, sorts documents, flags possible gaps and combines files. Its simplifying assumptions reveal what a more dependable review workflow needs.
This case study draws on the project repository and available build history. Problem framing and proposed improvements are editorial synthesis; unverified results are not presented as measured outcomes.
Some repositories are private; access is required to view those references.