LLMOps-Driven Robotic Process Automation Approach

摘要

The growing reliance on digitized documentation across industries has amplified the need for intelligent systems capable of extracting and structuring data from unstructured visual inputs, such as scanned invoices. Traditional Optical Character Recognition (OCR) systems, though effective in extracting raw text, often produce fragmented, inconsistent, or incomplete output that hinders downstream processing tasks. To address this challenge, we present an MLOps-Integrated Robotic Process Automation (RPA) pipeline that leverages a fine-tuned lightweight Large Language Model (LLM), specifically ‘unsloth/mistral-3B-v0.3-bnb-4bit‘, to convert noisy OCR text into clean, structured JSON formats suitable for database ingestion. The proposed system continuously monitors a target folder, validates image inputs, applies OCR, and utilizes the LLM to generate structured data, which is validated and stored. Crucially, the system integrates a feedback loop: after every 500 validated records, the model is fine-tuned and evaluated u sing metrics such a s B LEU Score, Levenshtein Distance, F1-Score, and Mean Squared Error (MSE). The best-performing version is auto-deployed using an MLOps workflow. Experimental e valuations demonstrate a B LEU score improvement exceeding 80% over the base model and substantial gains in sequence similarity, with minimal trade-offs in F1 stability and MSE. This research highlights a scalable and cost-efficient alternative to large-scale LLMs, showing that targeted fine-tuning within an automated lifecycle can significantly enhance OCR post-processing accuracy and utility.