AI Business Invoice Extractor

Manually re-typing data from stacks of invoices is slow, error-prone work. The AI Invoice Extractor solves this: drop in up to 100 PDF invoices, and the app reads them right in your browser, extracts just the fields you care about using an LLM, and hands you back a ready-to-use Excel spreadsheet. Everything runs without a backend — raw invoices never leave the user's machine unless they explicitly opt into cloud storage.

Designed and built the full application end-to-end: the browser-based OCR pipeline, the AWS Lambda AI extraction service behind API Gateway, and the Excel assembly and export flow.

• A drag-and-drop uploader that takes up to 100 PDFs at a time and automatically skips duplicates.
• Custom column selection so users decide exactly which fields to pull out of their invoices — invoice number, date, vendor, amount, anything.
• One-click extraction that turns a folder of PDFs into a single structured spreadsheet.
• Live progress tracking so users always know where their batch is, with clear error messages when something fails.

The key decision was local-first processing: OCR runs entirely in the browser, and only lightweight extracted text travels to AWS Lambda for LLM structuring — eliminating upload bottlenecks, cutting costs, and keeping sensitive documents on the user's device.

• Built with React 19, TypeScript, Vite, and Tailwind CSS for the interface.
• Tesseract.js and pdfjs-dist run OCR in a web worker, so PDF rendering never freezes the UI and files never leave the browser.
• AWS Lambda behind API Gateway calls an LLM to turn raw OCR text into structured rows, batched 20 invoices at a time to stay within Lambda's payload and timeout limits.
• A Rust-compiled WASM module decodes large base64 payloads roughly 10× faster than JavaScript during Excel export.
• Excel workbooks are assembled client-side with SheetJS from lightweight JSON responses instead of heavy file downloads.
• An optional S3 presigned-upload path exists for server-side processing and degrades gracefully with a warning banner when unconfigured.
• React Compiler provides automatic memoization to keep the interface responsive during long-running batches.

• Deployed live on Vercel as a fully static app with zero backend infrastructure for core processing.
• Raw invoices never leave the browser unless the optional S3 path is explicitly enabled — privacy by default.
• Batch processing handles 100-file workloads reliably, with per-batch failure isolation and progress reporting.