Layout-Preserving Document Synthesis and LLM-Assisted Benchmarking for Low-Resource OCR: A Case Study on Nepali

Document OCR for low-resource languages written in complex scripts, such as Nepali (Devanagari), faces an acute shortage of annotated, layout-diverse datasets. To address this bottleneck, we introduce an automated, layout-preserving document synthesis pipeline that converts high-resource datasets (SROIE, FUNSD, and XFUND) into Devanagari using LaMa image inpainting and aspect-ratio-aware text rendering. Additionally, we propose a semi-automated benchmarking workflow utilizing spatial grid-packing and a Multimodal Large Language Model (Gemini 3.8 Flash) to filter layout noise and refine ground-truth annotations. An expert human audit over 5,000 candidate crops demonstrates that our MLLM verification pipeline improves evaluation benchmark accuracy from 90.20% to 95.35% with zero observed false-positive noise flags across the audited sample, while reducing manual annotation labor by over 88%. Evaluating this framework with lightweight models, fine-tuning a PP-OCRv6 Small detector on our synthetic data improves word-level H-mean from 0.06 to 0.77, while recognizer fine-tuning pushes word-level accuracy from 84.56% to 91.73%, outperforming existing open-source baselines across both tasks. Our fully fine-tuned end-to-end pipeline delivers an ultra-lightweight, production-ready system (< 20MB footprint, < 60ms GPU latency).