ledge

Purpose

Small businesses and finance teams deal with invoices that arrive in every form imaginable, photographs of paper receipts, scanned PDFs, email attachments, and hand-filled forms. Processing these manually is slow, error-prone, and produces no structured record for future analysis. Existing general-purpose AI tools can read a single document, but they don't maintain context across a batch, can't trigger downstream actions, and produce no audit trail.

This project builds a multi-modal AI pipeline on Google Cloud that ingests fragmented invoice and procurement documents regardless of format or quality extracts structured information using Gemini 1.5 Pro and Cloud Vision API, makes a context-aware decision (valid, duplicate, anomalous, or flagged for review), explains which field or region of the document drove that decision, and logs every input-output pair to BigQuery for financial analytics and compliance.

Goals

• Accept documents in any combination of formats (image, PDF, plain text) through a single ingestion endpoint backed by Cloud Storage
• Use Gemini 1.5 Pro via Vertex AI to extract, interpret, and cross-reference document fields without requiring clean or complete inputs
• Maintain session and batch context so that a duplicate invoice submitted ten minutes after the original is caught, not processed twice
• Return a structured decision with a human-readable explanation citing the specific evidence from the document
• Log every transaction immutably to BigQuery, enabling downstream spend analytics, vendor patterns, and rejection rate tracking