Compositional AI-Service Pipeline to Generate Interactive Structured-Data from Scanned Images

This paper presents a compositional Artificial Intelligence (AI) service pipeline for generating interactive structured data from raw scanned images. Unlike conventional document digitization approaches, which primarily emphasize optical character recognition (OCR) or static metadata extraction, the proposed framework adopts a modular architecture that decomposes the problem into specialized AI services and orchestrates them to achieve higher-level functionality. The pipeline integrates core services including OCR for text conversion, image recognition for embedded visual content, interactive form modelling for structured data and NLP for extraction of structured representations from raw text. The form models incorporate various rules like value-type filtering and domain-aware constraints, thereby enabling normalization and disambiguation across heterogeneous document sources. A key contribution is the interactive browser linking extracted structures back to the original scanned images, thus facilitating bidirectional navigation between unstructured input and structured content. This functionality enhances interpretability, supports error analysis, and preserves the provenance of extracted information. Furthermore, the compositional design allows each service to be independently optimized, replaced, or extended, ensuring scalability and adaptability to diverse application domains such as press archives, enterprise repositories and government documents.