AI Expense Automation Platform

Finance teams often receive invoices and receipts through multiple channels, requiring personnel to manually review each document, extract key information, verify its accuracy, and encode the data into internal systems. This repetitive process consumes valuable time and increases the likelihood of human error.

As document volumes grow, manual processing becomes increasingly difficult to scale. Inconsistent document formats, handwritten receipts, and varying vendor layouts further complicate extraction and validation, creating delays in approval and reimbursement workflows.

The objective of this project was to design a secure, self-hosted automation platform capable of receiving documents, extracting structured information using OCR and AI, validating the results, storing records in a centralized database, and synchronizing approved data with business applications—while minimizing manual intervention.

I designed and implemented a production-ready, self-hosted AI automation platform that streamlines the entire expense submission and approval process. Rather than relying on multiple disconnected applications, the solution integrates messaging, workflow automation, OCR, artificial intelligence, database storage, and reporting into a single automated pipeline.

Expense documents are submitted through a Telegram Bot, which serves as the primary user interface for remote employees. Every upload triggers an n8n workflow responsible for orchestrating the complete processing lifecycle, including document validation, OCR extraction, AI processing, business rule validation, database updates, and application integration.

Document images and PDFs are processed by a custom OCR microservice built with FastAPI, Tesseract OCR, Poppler, and ImageMagick. The service performs image enhancement, PDF conversion, and optical character recognition before forwarding structured text to the AI layer for intelligent interpretation.

Artificial intelligence is provided through Ollama running locally for privacy-sensitive workloads and OpenRouter for cloud-hosted large language models when additional reasoning capabilities are required. The AI engine extracts business entities such as vendor information, invoice numbers, billing periods, invoice amounts, account codes, and document categories while handling inconsistent invoice formats from multiple suppliers.

Processed records are stored in PostgreSQL, creating a centralized and searchable repository of expense transactions. The workflow automatically synchronizes validated data with AppSheet, providing finance personnel with a familiar interface for review, approval, reporting, and audit tracking.

The entire platform is containerized using Docker and deployed on an Ubuntu Server. Cloudflare Tunnel provides secure HTTPS access without exposing internal services directly to the public internet, while Tailscale enables secure administrative access from any location using VS Code Remote SSH.