← All projects

My thesis

Connected Intelligence

A world where fragmented data becomes shared reasoning.

Point of viewArchitecture

The thesis

Most organisations don't lack data or models — they lack connection. Documents, systems, people and reports each hold part of the answer, in different formats and vocabularies. Intelligence emerges when they share context, can be reasoned over together, and the answer lands where the decision is made.

Why now

Language models made reasoning over messy information cheap — and made it just as easy to build confident systems on inconsistent foundations. The bottleneck has moved from “can a model answer?” to “can we trust the answer, and does it reach the person who needs it?”

From sources to decisions, in five layers.

feedback: usage · corrections · gapsSOURCESDocuments& policiesSystems& databasesWork trackers(JIRA, Monday)Reports & BIExternal news& dataINGEST & QUALITYConnectors& pipelinesValidation& DQ checksEntity resolution& de-dupLineage& metadataSEMANTIC LAYERKnowledge graph(ontology)Vector index(embeddings)Business glossary& metricsREASONINGRetrieval:GraphRAG + vectorText-to-SQLAgents & toolsEvaluation& guardrailsACTIONSearch & Q&AEmbedded inworkflows (PMT)Briefs, profiles& alertsAPIsGOVERNANCE& TRUSTEntitlement-aware accessProvenance & citationsAudit trailHuman-in-the-loop123445
Where my work sitsDeep dive on the governance & trust layer: Faithful by Design →

The nuances that make it work.

01

Connect, don't centralise

A semantic layer over systems where they live — not another copy in another lake.

02

Meaning before models

Most wrong answers are definitional. Governing the ontology drove our 40%+ accuracy gain.

03

Graphs + vectors

Vectors find what's similar; graphs know what's related and why. Real questions need both.

04

Trust is a feature

Show sources, respect who's asking, measure quality. Unverifiable answers don't get used.

05

Meet people where they work

Match Studio mattered once it lived inside PMT — at the moment of decision.

06

Experts curate, not answer

Put SME knowledge in the ontology once. That's how SME dependency fell 60%.

07

Every question teaches

Usage, corrections and unanswered questions feed back — so the system compounds.

What connected intelligence is not.

A chatbot on a data lake

Fluent answers on undefined data.

One model to rule them all

A better model won't fix bad definitions.

Governance as a final gate

Bolted on late, it blocks launches.

Five steps from fragmented to connected.

Level 1FragmentedAnswers depend on who you know
Level 2IntegratedShared platforms, different definitions
Level 3UnderstoodGoverned semantic layer
Level 4ReasonedAI answers with sources & access control
Level 5ConnectedEmbedded in workflows, always learning
What I'm building toward

Asking a question of the organisation should feel like asking a colleague who has read everything — and can show their working.

Next projectFaithful by Design
→