Book a demo

Technology · how it works

From client data
to agents and apps

Every layer reshapes the one beneath it. The environment turns raw records into a shared model; agents and applications work on top of it.

The approach

Map · Build · Activate

1/3

The systems the client already runs — ERP, CRM, documents, public registries — come in without a migration. The model is built on top of them, not in their place.

Scroll to compose the stack
INGESTIONOBJECT / ORDERTYPEorderLINKS12SOURCEERPAGENTCHATHello, world!FIELD APPMOBILEOPS APPWEBCREATE ORDERERPCRMTICKETINGWMSMESPIME-MAILFILES
Agent
Environment
Agents & apps
Agents and applications that act on the model
They read and write the same objects
View Engine
Views that expose the model to any surface
Decides what each surface can read
Ontologies
Entities and relationships that give data meaning
Objects, links and properties in a single model
Workflow
Pipelines that clean, merge and route records
Every transformation step, orchestrated
Data sources
Client databases, files and third-party systems
Records normalised into single entities

Four engines

What goes in, and what comes out

Seen from the side the stack is a pipeline. Seen from above it is a ring: four engines at the centre, and for each one a surface the client actually uses.

01Customer Model

Ontology engine

Gives data meaning before anyone queries it.

Dynamic entity types, computed fields and unified relationships between entities: semantic, referential, structural. Not a fixed schema decided up front, but a model the client extends while working.

  • Semantic · Kinetic · Dynamic
  • Human + AI Decision Model
  • Data · Logic · Actions · OSDK
  • Documents, images, video, geodata, audio
02Customer's Workflow

Workflow engine

Brings in the data that already exists, and keeps it current.

Scheduled, event-driven or API-driven automations, on top of a connection registry. Records arrive from the client's systems without a migration: the model is built above them, not in their place.

  • Scheduled · event-driven · API automations
  • OSINT, ERP, MES and proprietary-source connectors
  • Transactional, IoT, geospatial, relational, unstructured
03Customer's Application View

View engine

Decides what each surface is allowed to read.

A dashboard engine and a catalogue of composable widgets: entity lists, KPIs, aggregates, maps, three-column detail views. Views are versioned, so a configuration is an object with a history, not a browser state.

  • Dashboard engine · widget catalogue
  • Entity views and detail views
  • Versioning and time machine
04Customer's Prompt

Agent lifecycle

Agents that work on the resolved model, not on raw documents.

Building, orchestration and evaluation suites. The agent is not handed a collection of domain tools but two universal surfaces, and that is the point: what it can do is exactly what the user who invoked it can do.

  • Agent building · orchestration · evaluation
  • Commercial, open-source or self-hosted LLMs
  • PII moderation, caching, rate limiting

The pipeline

How a record becomes intelligence

  1. 01PULLPer-tenant query fan-out
  2. 02PRESELECTDiscarding what is not relevant
  3. 03SCRAPEFetching the content
  4. 04CONTENT-ANALYSISReading and classification
  5. 05EMBEDVectors for semantic search
  6. 06DIFFDeduplication against what already exists
  7. 07MC-EXTRACTExtraction of observed facts
  8. 08EXPANDFollowing the thread deeper
  9. 09STORY-PROCESSSynthesis into a traceable story

A fact lives in exactly one layer

This is the rule that holds the model together. Intelligence enters as a FACT; from there it is consolidated into a profile or a relationship, kept as a signal, or rejected. It is never copied by hand from one layer to the next — that is how provenance survives.

Profile
An entity's current attributes · User or consolidation
Relationships
Every link between entities · User, consolidation or enrichment
Facts
Intelligence observed in documents · Pipeline or manual entry

Governance

Why it holds up under scrutiny

Who sees what

Visibility at four levels: role, direct assignment, group, and cascade from the linked entity. Resolution walks down the group tree, so a permission granted once at the top need not be repeated below.

What an agent can do

Exactly what the user who invoked it can do: the agent works with a token in that person's name, not with a system key. There is no separate perimeter to keep in sync, because there is no second perimeter.

What stays on the record

Append-only audit trail, per-tenant isolation, approvals and checkpoints on the steps that matter. Every claim in an output traces back to the document it was observed in.

What happens if you say yes

Three steps

01Onboarding

Connect

The sources that already exist come in without a migration and without a new data warehouse. Nobody moves anything to begin — that is the condition that makes the first meeting a decision rather than a project.

02Practice

Compose

The model extends while you work: entities, relationships, views, automations. Agents are built on top of the resolved model, with declared tasks and evaluation suites, not with free-form prompts.

03Governance

Operate

Four-level visibility cascading down the group tree, permissions with no shortcuts on names, an append-only audit trail, per-tenant isolation. It is the longest of the three steps, and it is the one that holds up under scrutiny.

Half an hour, one real source, and a model built live.

Book a demo