Handwritten Document OCR Platform

Old handwritten manuscripts present significant legibility challenges due to ink fading, paper degradation, and non standard writing styles. Standard optical character recognition engines failed to transcribe these images accurately. We needed a complete cloud infrastructure to handle document uploads via Cloud Storage, track processing status through Cloud Firestore, store results, and return structured API responses. We initially planned to use Gemini 3.1 Pro through OpenRouter, but quickly encountered severe API rate limits on our new developer account during batch processing runs. We also needed to comply with strict General Data Protection Regulation guidelines since the manuscripts contained historical records with personal names and locations. Sending high resolution images to multi modal models consumed a large number of input tokens, which threatened to make the project financially unviable.

We built a complete serverless pipeline on Google Cloud Platform. The FastAPI application runs on Cloud Run with Cloud Load Balancing handling SSL termination and health checks. When a client submits a document, the image is uploaded to a GCS input bucket in the europe-west region. The application creates a job record in Cloud Firestore with an initial status of pending, then begins processing. The document is binarized, margins are removed, and the content is sliced into smaller tiles to optimize token consumption. Each tile is sent to Gemini 2.5 Flash on Vertex AI in the Europe West region. If Vertex AI returns a rate limit error, the router falls back to Gemini 3.1 Pro via OpenRouter. As processing progresses, the Firestore job status updates from processing to complete or failed. The final transcribed text is written as a .txt file to the GCS output bucket, organized by date and job identifier. Cloud Monitoring collects custom metrics for API latency, token usage, and error rates.

This platform transcribes historical handwritten manuscripts using vision language models. It uses Python FastAPI for the API layer and combines Gemini 3.1 Pro via OpenRouter and Gemini 2.5 Flash via Google Vertex AI. Document images are uploaded to a Google Cloud Storage input bucket, processed through the OCR pipeline, and the transcribed text is stored as .txt files in a separate GCS output bucket. Cloud Firestore maintains real-time status tracking for each processing job with states like pending, processing, complete, and failed.

The production implementation successfully processed over one million manuscript pages. By switching from Gemini 3.1 Pro to Gemini 2.5 Flash and implementing token optimization, API costs were reduced by eighty five percent while retaining high transcription accuracy. Running on Cloud Run serverless infrastructure allowed the system to scale down to zero when idle, keeping operation costs minimal while maintaining prompt responsiveness.