Document Intelligence Platform

Part of the worlds-biggest-software-project initiative.

An AI-native, open-source platform for extracting structured data from any document type -- replacing expensive per-page APIs with zero-shot schema inference and agentic workflows.

Document Intelligence Platform is a self-hostable document processing engine that combines classical computer vision with multimodal foundation models to parse, classify, extract, and validate data from PDFs, images, office documents, and scanned forms. It is built for operations teams drowning in manual data entry, ML engineers feeding RAG pipelines, and IT architects who need auditable extraction without hyperscaler lock-in.

Why Document Intelligence Platform?

Per-page pricing is unsustainable at scale. Azure Document Intelligence, Google Document AI, and AWS Textract charge $1--$30 per 1,000 pages depending on features used. Costs become unpredictable at enterprise volume, and every additional API feature multiplies the bill.
Hyperscaler lock-in limits deployment options. All three major cloud offerings (Azure, Google, AWS) are proprietary and cannot be self-hosted. Regulated industries requiring data sovereignty have no viable path with these vendors.
Existing tools process documents in isolation. No incumbent platform natively reconciles data across related documents -- matching a purchase order to an invoice to a goods receipt still requires custom integration.
Template-based extraction cannot handle the long tail. Most platforms require upfront template definition or thousands of labelled examples before processing a new document type. Novel layouts and edge cases are left unaddressed.
Extraction decisions are opaque. Users across the industry report that low-confidence fields come with no explanation, forcing reviewers to re-read entire documents rather than validating specific extraction rationale.

Key Features

Document Ingestion and OCR

Multi-format ingestion: PDF (native and scanned), DOCX, PPTX, images (JPEG, PNG, TIFF), with expansion to XLSX, HTML, EML, MSG, and EPUB
OCR pipeline combining classical computer vision with Vision Language Model refinement for high accuracy on degraded inputs (faxes, low-resolution scans, stamps)
Multi-language support targeting 50+ languages at MVP, 100+ at v1.1

Structured Extraction and Schema Inference

User-defined JSON schema for configurable field extraction
Zero-shot schema inference: the platform infers extraction schemas directly from novel documents without labelling or template definition
Confidence scores per extracted field with natural-language explanations for low-confidence results
Instruction-driven extraction using natural-language prompts to control output format and preprocessing

Classification, Splitting, and Routing

Document classification and type detection for automated routing
Composite document splitting: separate multi-type packages into single-class segments
Confidence-based routing to human review queues or automated downstream actions

Human-in-the-Loop Review

Annotation, correction, and feedback UI for field-level review
Reviewer corrections feed back into model improvement via few-shot in-context learning
Natural-language explanations accelerate review by surfacing extraction rationale rather than raw confidence numbers

Cross-Document Reconciliation and Agentic Workflows

Match fields across related documents (purchase order to invoice to goods receipt to payment confirmation) and flag discrepancies automatically
Agentic workflow chaining: extraction, validation, enrichment, and downstream action triggers (ERP write-back, approval routing, exception escalation) without external orchestration
Async batch processing API (REST) with webhook notifications on completion

Connectors and Integration

REST API with SDKs for Python and JavaScript (additional languages planned)
Connector library for S3, Azure Blob, GCS, SharePoint, and major vector databases
MCP server implementation for integration with AI agent frameworks
Domain-specific pre-built extractors for invoices, receipts, identity documents, and contracts (v1.1)

AI-Native Advantage

Multimodal foundation models (GPT-4o, Gemini, Claude) can extract structured data from novel document layouts without task-specific training data, eliminating the weeks of labelling that classical ML pipelines demand. The platform uses these models for adaptive schema inference -- determining what to extract directly from the document itself -- and for generating natural-language explanations of low-confidence fields, enabling reviewers to validate results in seconds rather than minutes. Cross-document reasoning, where an LLM reconciles data across related documents and surfaces discrepancies, is a capability that rule-based and single-model IDP tools fundamentally cannot provide.

Tech Stack & Deployment

The platform supports self-hosted, cloud, and hybrid deployment modes. The OCR pipeline combines open-source engines (producing hOCR and ALTO XML output) with VLM-based refinement for edge cases. Document archival targets PDF/A (ISO 19005) compliance for regulated industries, and extraction results can be represented using the W3C Web Annotation Data Model for portability. Integration with ECM repositories (SharePoint, Documentum, Alfresco) is supported via CMIS (OASIS). The architecture separates the extraction engine from the review UI and workflow orchestrator, allowing each component to scale independently.

Market Context

The intelligent document processing (IDP) market was valued at approximately $1.8B in 2024 and is projected to reach $9.5B by 2030 at a ~30% CAGR, driven by enterprise automation demand and LLM adoption lowering the barrier to high-accuracy extraction. Primary buyers include operations leaders in financial services, insurance, healthcare, and logistics; IT architects embedding document understanding into ERP and CRM workflows; and developers building RAG pipelines that require reliable structured data from unstructured documents. Incumbent pricing ranges from per-page fees ($1--$30 per 1,000 pages for hyperscalers) to flat-rate SaaS ($500--$2,000+/month for vertical specialists like Docsumo and Rossum) to six-figure enterprise deals (ABBYY, Hyperscience), leaving significant room for an open-source alternative with predictable costs.

Project Status

This project is in the research and specification phase.
Contributions, feedback, and domain expertise are welcome.

Contributing

We welcome contributions from developers, domain experts, and potential users. See CONTRIBUTING.md for guidelines.

Important: All contributions must be your own original work or clearly attributed open-source material with a compatible licence. Copyright infringement and licence violations will not be tolerated and will result in immediate removal of the offending contribution. If you are unsure whether a piece of code, text, or other material is safe to contribute, open an issue and ask before submitting.

Licence

Licence to be determined. See discussion for context.