How SemanticTwin.ai builds your semantic model

Bring what you already have

SemanticTwin accepts 46 kinds of input, in seven groups.

Each file is checked first. It gets a fitness score, with strengths, weaknesses and a recommended use, to show how much it can contribute.

About code: SemanticTwin reads the code that defines your data: schemas, data models and API definitions.

Requirements and processes 9 input types

Business Requirement Document (BRD)
Functional Requirement Document (FRD, also called FSD)
Standard Operating Procedure (SOP)
Operating Model Document
BPMN Diagram
Flowchart
Swimlane Diagram
UX / User Journey Flow
Service Blueprint

Rules and policies 9 input types

Policy Manual
Risk or Compliance Rules
Eligibility Rules
Pricing Rules
Regulatory Guidance
Approval Matrix
Decision Tables
Audit Requirements
Legal Definitions

Glossaries and data models 6 input types

Enterprise Business Glossary
Departmental Glossary
Business Data Dictionary
ER Diagram
Conceptual Data Model
Logical Data Model

APIs, services and events 9 input types

OpenAPI / Swagger Specification
GraphQL Schema
WSDL Service Definition
XML Schema Definition (XSD)
Avro Schema
API Behaviour Description
API Data Access
Event Documentation
Webhook Descriptions

Databases and code 6 input types

SQL Database Schema (DDL)
Python ORM Model
Database Profile Report
Database Export (CSV)
Parquet / Data Lake Files
Tabular Data File

Analytics and data management 5 input types

Analytics Metric Catalog
KPI / Metric Definitions
Reporting Field Definitions
Data Lineage Documents
Data Quality Reports

Knowledge and systems 2 input types

Wiki or Confluence Pages
Source System Descriptions

The journey, step by step

Six steps in two lanes. SemanticTwin does the reading, matching and drafting; your team makes the decisions.

Your documents and code

Requirements (BRD, FSD), process diagrams, API docs, database schemas (DDL) and code

SemanticTwin does the work
Read everything

It reads the text, the tables and the diagrams inside documents. Structured files such as database schemas and API specifications are read exactly by parsers, not guessed. Prose is read by AI.
SemanticTwin does the work
Cross-check and discover

The same business thing often appears under different names in different files. SemanticTwin matches them up and flags where sources disagree. It finds the business things, processes and rules three separate times, and shows how consistently each one appeared.
Your team makes the decisions
Confirm what’s in scope

People decide what belongs in the model, what doesn’t, and what is only for reference. Capabilities, processes, objects and rules are each marked In scope, Out of scope or Reference only.
SemanticTwin does the work
Build the model, step by step

The model is built in 16 steps across three phases: foundation, core modelling and formalisation, with proof on every fact.
Your team makes the decisions
Review at each gate

The build pauses at review gates, so your team can correct anything before the next step builds on it.
Your team makes the decisions
Approve and publish

A model goes live only after it has been reviewed and approved.

A trusted semantic model of your system

Inside the build

The model is built in 16 steps across 3 phases, with 4 review gates along the way.

Three phases

Phase 1
Foundation
Phase 2
Core Modelling
Phase 3
Formalisation

Four review gates

When the build pauses at a gate, your team can correct anything before the next step builds on it.

Discovery & Triage
Blueprints
Model Review
Publish Review

Two run modes

Checkpoint

Pauses at all four gates.

Full auto

Approves the Discovery & Triage and Model Review gates automatically. Blueprints and Publish Review still wait for your team.

Three quality tiers

Exploration

Fastest

Analysis

Balanced

Publication

Highest quality

Trust and proof

Every fact carries its proof, and people outrank software.

Proof for every fact

Every element carries a quote and the file it came from, and quotes are checked word for word against the source.

A quote that can’t be confirmed is marked unvalidated, never hidden.

Confidence labels

Discovery runs three times, independently. Each item is labelled by how many of the runs found it.

ExplicitFound in 3 of 3 runs
ImpliedFound in 2 of 3 runs
WeakFound in 1 of 3 runs

People outrank software

Facts from structured sources rank above facts inferred from prose. Your team’s decisions rank highest, and a re-run never overwrites them.

Your team’s decisions
Facts from structured sources
Facts inferred from prose

Explore it in the Semantic Playground

Explore the model visually: the business things, what each one means, and where it is stored.

New developers learn the system there instead of piecing it together from the code. New analysts, testers and support staff learn how the business works.

Four views

Concept Map
Knowledge Graph
Table
Dashboard

Five panels

Inspect
Ask
What-if
Issues
History

Ask answers plain-English questions, only from the model.
What-if simulates a change, so you can see which findings and rules it affects.
History records every change: who made it, when and why.

What you get

A published model, twelve documents, exports and an MCP server for AI assistants.

The published semantic model

Business things (concepts)
Their details (attributes)
Connections (relationships)
Rules
Life stages (states and transitions)
Links to real tables and columns
Proof

Twelve generated documents

Domain Overview
Business Glossary
Process Flows
Data Dictionary
Entity-Relationship Diagram
Relationship Matrix
PII Register
Business Rules Catalog
Validation Report
Traceability Matrix
Design Patterns
State Lifecycle Diagrams

Exports

The model in Turtle, JSON-LD, RDF/XML and N-Triples
Documents as Markdown or PDF
An export format designed for spec-to-code pipelines

An MCP server for published models

MCP (Model Context Protocol) is the standard way to plug AI assistants into business systems. SemanticTwin includes an MCP server for published models. It lets AI assistants:

list models and search concepts
read a concept
get query guidance
find the join path between two concepts
see the evidence behind a fact
run read-only queries checked against the model’s rules

Teams can also build their own MCP tools on top of the model.

Personal-data detection

SemanticTwin detects fields that hold personal information and lists them in a PII register.

An agent-readiness score

A score for how ready a model is for AI agents, with a list of gaps to fix first. It’s available through the API only.

From draft to published

Draft
Submitted for review
Approved
Published

Every change records who made it, when and why.

A worked example: BAI procurement

The BAI procurement system is a purchasing platform for building materials. A buyer raises a request for materials, invited sellers quote, both sides negotiate, and the accepted quote becomes a purchase order.

The system has 11 database tables, and their names mean little to anyone outside the team that built it. SemanticTwin was given two files: the requirements document and the database schema. From them it worked out what each table means.

12 business things
69 details
18 connections

Each linked to the sentence or schema line it came from.

What each database table means to the business
In the database	What it means to the business
`prch_ordr`	A material request raised by a buyer
`prch_ordr_split`	A group of lines, split out so different sellers can quote
`prch_ordr_offers`	A seller’s quote
`prch_ordr_offers_hist`	One round of negotiation on a quote
`mvt_prch_ordr`	The purchase order created when a quote is accepted
`prch_ordr_trx_hist`	The audit trail of every status change

Status codes, decoded

VASSA vendor has been assigned
ODCFThe order is confirmed
ODDPThe order has been dispatched

A real rule

Once a seller submits a quote, its prices can’t be edited.

The analyst’s question, now with a clear path

“Which material requests turned into purchase orders last month?”

Every purchase order points back to its material request and to the quote that won. A person or an AI agent can follow that path without first spending days decoding table names.