From Manual Chaos to Automated Reimbursements

Overview

We built this reimbursement system to replace a manual process that depended on email threads, file attachments, and repetitive review work. The platform handles receipt intake, extraction, validation, and reporting in one pipeline, while keeping humans in control of final review where needed.

Problem

Before this system, departments submitted receipts in inconsistent formats and admins had to process everything by hand. Data quality varied from one submission to another, and monthly reporting required too much manual consolidation. The team spent time fixing format issues instead of reviewing actual reimbursement decisions.

What We Built

The core of the solution is a staged workflow: upload, clean, extract, validate, store, and report. OCR and document preprocessing clean up noisy files before extraction starts, which improved consistency for downstream steps.

For extraction, we used local LLMs and avoided large all-in-one prompts. Documents are split into smaller chunks and processed in parts. This reduced extraction errors and made the output more predictable.

Reporting is generated from the same structured data pipeline, so monthly summaries and user-level reimbursement reports no longer require manual assembly. Managers also get a dashboard view of totals and pending versus approved requests.

Stack

The frontend uses Next.js with Tailwind CSS and shadcn/ui. The processing layer includes OCR and document cleaning, and the AI extraction layer uses self-hosted local LLMs.

Outcome

Most of the repetitive admin work was removed from day-to-day reimbursement handling. Processing and reporting cycles became faster, and the output quality became more consistent because every submission now passes through the same steps.

Live Project

https://reimburse.app