The backbone for UK AI
Any AI algorithm searching for financial signals, industry context, locations or training data ends up reinventing the same fragmented integration. The Foundry replaces that effort with a single connected backbone — so models train, retrieve and predict on grounded, provenance-tracked UK facts instead of hallucinating.
See the data sourcesThe scale advantage
The combination is the product. Any one of these sources is useful; all of them, joined on one identity layer with provenance, is transformational for AI.
40+
connected data sources
Corporate, financial, procurement, ESG, property, workforce, trade and derived signals — joined on a single canonical UK entity spine.
17.2M
resolved UK entities
Companies, directors, owners and connected places — the grounding any UK-focused AI system needs to stop guessing.
1
knowledge graph
One unified graph instead of 40 disconnected feeds, so models retrieve consistent, provenance-tracked facts.
0
data brokers to integrate
Replace months of vendor procurement, schema mapping and entity reconciliation with a single sovereign backbone.
How it works
Click any stage to see what flows in, what comes out and why it matters for AI systems built on the Foundry.
Stage detail
Authoritative public, licensed and proprietary UK datasets are ingested on stable cadences — corporate filings, financials, procurement, ESG, property, workforce, trade and derived signals.
Inputs
Outputs
Example payload — Raw sources
One raw record arriving from Companies House
entity: SDF_03977902
// source: companies_house fetched_at: 2026-05-09T03:14:11Z
{
"CompanyNumber": "03977902",
"CompanyName": "EXAMPLE HOLDINGS LIMITED",
"RegAddress.PostCode": "EC2A 4NE",
"CompanyStatus": "Active",
"SICCode.SicText_1": "62012 - Business and domestic software development",
"IncorporationDate": "2000-04-12"
}Department view — Raw sources
Switch perspective to see what this stage delivers for each part of government and the SME ecosystem.
Selected perspective
Procurement
Stage: Raw sources
Tailored benefit
One ingest of every UK contracts, awards and supplier filings feed — no per-portal monitoring or per-vendor licence.
Use case stories
Showing stage: Raw sources.
The hallucination problem
Hallucinations are not a model failure first — they are a data failure. Models guess when grounding is missing, contradictory or unverifiable.
Foundation models trained on the open web have shallow, outdated and inconsistent knowledge of UK firms, contracts, ownership and places. They confidently fill the gaps with plausible fiction.
Even when teams stitch vendors together, identities don't match across systems. The model is fed contradictions, so it picks one — or invents a third.
Without source, timestamp and confidence on every fact, models cannot reason about freshness or trust. Stale data is treated as current truth.
Single-source signals miss the patterns that predict insolvency, supplier risk, growth or fraud. The model has no time series to reason over.
What the backbone delivers
The Foundry exposes its connected estate as the building blocks data-hungry AI systems actually consume.
Curated, deduplicated and licensed UK datasets formatted for supervised fine-tuning, instruction tuning and continued pre-training of domain models.
Drop-in retrieval over the connected knowledge graph so chat agents, copilots and decision systems answer with cited UK facts instead of plausible guesses.
Pre-engineered features for credit, supplier risk, growth, churn, fraud and demand models — built once on the sovereign spine, reused everywhere.
Stable, governed APIs that AI agents can call as tools — with consistent schemas, rate limits and audit trails suitable for production.
Use the same backbone that powers your model to evaluate it against a UK-specific benchmark suite — so accuracy claims are grounded in real sovereign data.
Every dataset arrives with licence terms, lineage and DPIA-ready notes — so AI systems built on the backbone are deployable in regulated and public-sector settings.
The stack
The same backbone serves training, retrieval, features, agents and evaluation — one investment, reused across every AI workload.
Layer 01
Source integration
40+ authoritative UK feeds harmonised into one ingest layer.
Layer 02
Entity resolution
Canonical identifiers across companies, directors, owners and places.
Layer 03
Knowledge graph
Connected facts that AI systems can traverse instead of guessing.
Layer 04
Provenance & versioning
Source, time and confidence on every assertion.
Layer 05
Feature & embedding store
Pre-computed features and vectors for ML, RAG and agents.
Layer 06
Governed access APIs
Stable endpoints, rate limits, audit trails and licence metadata.
Before and after
Without the backbone
With the backbone
Who it serves
Foundation model labs
Continued pre-training and fine-tuning on a coherent UK corpus instead of scraped fragments.
Enterprise AI teams
Build copilots for credit, procurement, sales and risk on grounded sovereign data.
Government AI projects
Deploy assistants and decision systems with provenance suitable for public-sector accountability.
Startups and SMEs
Ship vertical AI products without spending the first two years rebuilding a national data layer.
Developer interfaces
The same backbone that powers AI training is exposed through stable, documented endpoints — designed for production agents, copilots and analytics.
/v1/entities/{id}
Resolve a single canonical entity by SDF ID, Companies House number, domain or VAT number. Returns the unified record with per-field provenance.
GET /v1/entities/SDF_03977902
Authorization: Bearer <token>
200 OK
{
"id": "SDF_03977902",
"name": "Example Holdings Ltd",
"status": "active",
"registered_address": { "postcode": "EC2A 4NE", "lat": 51.522, "lon": -0.082 },
"_provenance": {
"name": { "source": "companies_house", "as_of": "2026-04-12" },
"status": { "source": "companies_house", "as_of": "2026-05-02", "confidence": 0.99 }
}
}/v1/search
Faceted search across entities, contracts, properties and people. Supports filters on geography, sector, financials, ownership and dates.
POST /v1/search
{
"type": "entity",
"filters": {
"postcode_area": "M",
"sic_section": "C",
"employees_gte": 10,
"status": "active"
},
"limit": 50
}/v1/graph/traverse
Walk relationships across the knowledge graph — owners, directors, contracts, suppliers, premises — with depth and edge-type controls.
POST /v1/graph/traverse
{
"start": "SDF_03977902",
"edges": ["owns", "controlled_by", "awarded"],
"depth": 3,
"return": ["entity", "contract"]
}/v1/changes?since={cursor}
Resumable, append-only stream of every change with source, timestamp and previous value. Drives live agent context, alerting and incremental retraining.
GET /v1/changes?since=2026-05-09T00:00:00Z
Accept: text/event-stream
event: entity.updated
data: { "id":"SDF_03977902", "field":"status",
"from":"active", "to":"in_administration",
"source":"insolvency_service", "as_of":"2026-05-10T08:14:00Z" }Authentication, governance and audit
Every interface is wrapped in identity, policy and lineage — so departments, enterprises and regulated workloads can trust both the data and the system that served it.
Short-lived OAuth 2.0 access tokens issued per workspace, with optional mTLS for government and enterprise tiers. Service accounts and human users carry distinct identities so audit trails are unambiguous.
Role-based access combined with dataset-level policies enforces who can read which sources and at what granularity. Licence terms travel with every response so downstream use stays compliant.
Every API call, schema change and data release is recorded against the calling identity, with the request, response shape and the dataset versions consulted — so any AI output can be traced back to its evidence base.