Platform
How Lintel is built.
A credible, non-confidential overview of the pipeline and the design choices that keep every output traceable.
Ingestion pipeline
From scan to verified obligation.
- 1
OCR
Scanned pages become text with layout.
- 2
Structure
Headings, tables and schedules are recognised.
- 3
Clause segmentation
Numbered clauses and cross-references are identified.
- 4
Extraction
Obligations and fields are proposed with citations.
- 5
Linking
Clauses, obligations and events are connected.
- 6
Verification
A person reviews low-confidence items.
Architecture
Layers
A simplified layered view. Interfaces marked planned are not yet available.
Design choices
What we insist on
Provenance on every field
Each extracted value stores its source page, text span and a confidence score, so any field can be traced and challenged.
Review queue
Low-confidence extractions are routed to a person. Approved values are versioned; edits are logged.
Hybrid retrieval, scoped by project
Keyword and semantic search run only inside the project’s own documents. Nothing crosses tenants.
Model-agnostic AI gateway
Model calls go through one gateway so models can be swapped or evaluated without changing product behaviour.
Evaluation on annotated sample contracts
Extraction is evaluated against annotated sample contracts before changes ship. We do not publish an accuracy percentage.
Multi-tenant isolation
Each customer’s data is separated logically, with role-based, project-scoped access and audit logging.
Evidence chain-of-custody
Every evidence item has a timestamp, a hash and an access log.
API and webhooks (planned)
A versioned REST API and webhook events are planned. They are not yet available.
Start with one contract.
Join early access, or book a walkthrough and we will show you how a single agreement becomes a live register.