Modern OCR in Resource-Constrained Environments

Optical Character Recognition (OCR) is crucial for automating scanned document processing, yet designing high-quality systems under resource constraints remains challenging. In this talk, we share insights from building OCR solutions for real-world applications. We trace the evolution of OCR pipelines from traditional character recognition approaches, which rely on dictionaries and rule-based post-processing, to modern solutions leveraging open-source frameworks, vision models, and LLMs for contextual correction, thereby enhancing extraction quality and robustness. Through concrete examples, we compare different approaches, discussing their strengths, limitations, and performance in practical scenarios. We present engineering trade-offs required to maintain efficiency, reliability, and maintainability under these conditions. Furthermore, we explore methods for combining visual and textual signals to extract structured information while ensuring a consistent and reliable user experience