Docupath is a document intelligence platform that transforms unstructured business documents - invoices, purchase orders, contracts, medical records, and more - into structured, actionable data. Unlike traditional systems that rely on a single model or template-based rules, Docupath uses an AI Model Garden architecture that orchestrates a network of specialised AI models across language, vision, layout, and domain reasoning. The platform is built around a human-in-the-loop principle: AI proposes, humans validate, and every decision is fully auditable.

Docupath's core architecture is a dynamic orchestration layer that analyses each incoming document, routes it to the appropriate subset of specialised models, and aggregates results through cross-model auditing. No single model handles everything — instead, the right combination is activated per document, per field, per context.

Crucially, the Model Garden is not limited to large language models. An orchestration algorithm evaluates each document and field and then selects the best tool for the job, whether that's a neural model or a deterministic algorithm. In practice, this means:

By combining learned models with deterministic computation, the Model Garden cross-checks results across methods, catching errors that any single approach might miss and raising confidence in the final extraction.

Every document moves through five stages:

Every extracted field links back to its source location in the original document. Reviewers can inspect the evidence, override AI decisions, and approve or reject documents through a structured workflow. All edits are logged with full before/after traceability.

For organizations processing high volumes, requiring manual review of every document can become a bottleneck. Auto Review lets Docupath approve repeat documents automatically: once a trading pair is trusted, matching documents that pass Auto Review's checks are approved without manual review, while any document that looks different is flagged and routed to a reviewer.

Auto Review is optional and configurable. Trust is built up per trading pair, and the identity, trust-ledger, and business-rule checks that gate auto-approval can be tuned per pair, so human-in-the-loop safeguards stay in place wherever oversight matters most. See the Auto Review article for the full workflow.

The platform combines vision models (for layout and handwriting), language models (for semantic meaning and context), and layout models (for structure-independent parsing) to handle the full range of real-world document quality and format variation.