Quoting
best4travel
Named client
Ocean-cruise quoting
- Quote time cut from 25 minutes to 3
- 95% of weekly cruise quote volume generated automatically
- In production since January 2026
In production since January 2026 · best4travel
Governed AI for regulated work. Your premises, your policy.
Shelby is an AI platform installed on hardware you control: models on your premises, a gateway of five controls on anything that leaves, and workload blueprints that arrive pre-governed. Proof in production since January 2026: best4travel cut cruise-quote time from 25 minutes to 3 and now generates 95% of weekly quote volume automatically.
- Quote time
- 25 → 3 min
- Volume automated
- 95%
- In production
- Jan 2026
In production
Four workloads in production. Five deployments.
Quoting · Reporting from operational data · Tender clarification and QA · DORA artefact production. Each reads messy inputs and produces the document a trained person would have produced, with a person checking where it matters.
Quoting
iqbranding
Named client
Branding and print quoting
- 50 reference test cases gate every release
Reporting from operational data
ARM
Internal production deployment
Advanced Radio Mapping, our in-house analytics practice, has measured exhibitor visitor traffic with on-site radio sensors since 2017. Shelby turns its sensor exports into the client report.
- Eight-section client report with visitor cohorts, for Palfinger in Las Vegas
- Second deployment at a European motorsport event, 2026 · delivered in .docx and .pdf
- Deduplication mismatch flagged before release
DORA artefact production
Irish-regulated insurer
Client under NDA
Register of Information and exit-plan build
- AI systems recorded on the Register of Information
- Accepted on a clean ESA validation pass
Tender clarification and QA
Amata Engineering
Internal production deployment
Electrical contracting, pre-award
- 150+ tender files read across two disciplines, ~90 of them drawings
- 250+ numbered clarifications consolidated; ~70 overlapping queries merged to ~50
- Two working days, estimator sign-off
For regulated firms, AI stalls at the governance boundary.
Your compliance function cannot approve an AI route it cannot see. Whether a given model version may process your data depends on the deployment, the endpoint and the contracting entity, and the answer changes every quarter.
Shelby arrives with the design, the governance manifest and the audit scope already written for each workload. Your risk function reviews a document, not a project. The second deployment costs less than the first.
Shelby Gateway
Five controls between your data and any model.
The Gateway runs on hardware you control. Every blueprint deploys on top of it.
Reversible pseudonymisation
An on-premises filter tokenises names, PPSNs, Eircodes, IBANs, account references and the classes you agree with us before any call leaves. Answers are re-identified on your side. If the filter fails, nothing goes out.
Independent egress audit
A second model, from a different model family and on separate hardware, inspects every outbound flow for residual identifiers and keeps a tamper-evident log of what it checked, caught and escalated.
Governed fine-tuning
Audit flags become training data only after a person labels them. Models are promoted with signatures, versions, rollback and training-data lineage.
Policy routing
Each workload class is routed by your policy, resolved against the Shelby Register: a dated evidence base of where each model version may lawfully be processed.
Verified private connectivity
Destination allowlists, private DNS, denied public egress and private endpoints, with routes verified continuously.
Where work may go
| Tier | Destination | Typical use |
|---|---|---|
| T0 | On-premises models, on your hardware | Raw data; highest sensitivity |
| T1 | EU-sovereign GPU capacity: EU-domiciled provider, EU facility | Pseudonymised personal data. The governed default |
| T2 | EU-region frontier deployments, each qualified on its documented residency | Pseudonymised work, where your policy accepts hyperscaler processing |
| T3 | Global endpoints | De-identified or non-sensitive work only |
Two substrates for the on-premises tier
- Our own stack — in production. NVIDIA hardware we specify, install and tune. Every deployment to date runs this way.
- Azure Local — in validation. Microsoft’s on-premises stack, run in a customer location or a partner-operated data centre, with disconnected operation supported. Microsoft offers it with the NVIDIA RTX PRO 6000 Blackwell Server Edition GPU — the same class we specify for production — and we are validating Shelby’s on-premises tier on it now. On either substrate the Gateway, keys and auditor run in your environment; public Azure is not a home for the Gateway, and Microsoft-hosted EU-region models remain a T2 route.
In a Microsoft estate
- Shipping today: sign-on through Entra; SharePoint and Fabric as document sources.
What Shelby is
- A platform, not a model. Language models running on your premises, with governed routes to external capacity when your policy allows it. Every route speaks the OpenAI-compatible API standard, so models run on your hardware or with any provider behind one interface.
- The Gateway. The five controls on this page, running on hardware you control, between your data and any model.
- Blueprints. A library of workload control baselines — quoting, reporting, tender QA, DORA artefacts — each arriving with its governance manifest and acceptance tests written.
- Deployed by AIfactor. We install it, tune it to your corpus and hand your compliance function the evidence. You operate it; overflow routes to governed EU capacity.
What Shelby is not
- Not anonymisation. Pseudonymised payloads are treated as personal data unless an assessment shows otherwise.
- Not your lawful basis, DPIA or transfer mechanism. It supports them; it does not replace them.
- Not a certified DLP guarantee. Detection is statistical; residual risk is measured and reported per entity class.
- Not a cloud service. The gateway, keys and auditor run in your environment.
Blueprints
Workloads that arrive pre-governed.
Each blueprint is a control baseline: inputs, outputs, where a person checks, what may leave and the tests that gate go-live. Your compliance function approves the use case.
In production
Batch and on demandQuotingPriced quotes from customer requests, supplier inventory and your price rules.Batch, per event or reporting periodReporting from operational dataClient-ready reports built from platform exports, with every figure traced to its source.On demand, per tenderTender clarification and QAPre-award clarification questions and QA checks, read from the full drawing set.Annual cycle and on changeDORA artefact productionThe Register of Information, exit plans and vendor packs, to ESA validation standard.
Ready to pilot
On demandLoan application handlingComplete, checked application files for your advisers.Batch and on demandInvoice processingSupplier invoices coded and matched, with exceptions flagged.BatchAccounts reconciliationBank and ledger matched, with breaks queued and an audit trail.BatchBilling cycleA draft billing run with anomalies flagged before release.
The 30-day pilot
Acceptance criteria fixed before kickoff.
The pilot ends against objective tests in the statement of work, evaluable by your compliance function as well as your engineers.
Offline pilot
No cloud contract. Your data never leaves the premises. Proves detection by entity class, re-identification integrity, the auditor and fail-closed controls, with routing simulated.
Governed-egress pilot
Adds approved external routes. Proves model quality, residency behaviour and private network paths against your own provider accounts.
Example acceptance criteria
- recall ≥ 0.95 on high-sensitivity identifiers
- F2 ≥ 0.90 aggregate
- zero egress of seeded canary identifiers
- fail-closed behaviour verified
- P95 latency and throughput measured
What a pilot delivers
Three use cases are named in the statement of work before anything is installed, and each is measured against your own current baseline — the time, cost and error rate of the workflow as it runs today — not against a vendor benchmark.
- Week 1
Install and connect. Hardware on premises, corpus indexed, access controls in place, and the baseline for each use case measured and signed off with your team.
- Weeks 2–3
The three use cases run alongside your existing process. Every output is scored against the acceptance criteria; discrepancies are logged and reviewed, not smoothed over.
- Week 4
The evidence pack is assembled and walked through: measured results against baseline, per-criterion pass or fail, audit-log extracts, and a production sizing recommendation.
You keep the evidence pack whether or not you proceed. It is written to be read by your compliance function, your auditor and your board — not only by engineers.
- Pilot
A pair of NVIDIA DGX Spark units on your premises, loaned for the pilot or purchased.
- Production
RTX PRO 6000 Blackwell-class servers running the same software, sized from pilot measurements — on your premises or in an Irish partner-operated data centre.
- Overflow
Work beyond your estate goes out pseudonymised, policy-routed and privately connected to governed EU capacity.
Fixed fee €25–50k, set by corpus scope. No hardware purchase required.
For partners
You host, connect and operate. We bring the blueprint, the pilot and the evidence pack.
Shelby fits managed service providers, network operators and data-centre operators who own the client relationship and the infrastructure it runs on.
Where it sits
On the client’s premises or in an Irish partner-operated data centre. The Gateway, keys and auditor stay in the client’s environment either way: the partner runs the racks and the connectivity; access, keys and the audit trail stay with the client.
Over your network
Private connectivity across your own network fits the verified-private-connectivity control: destination allowlists, private DNS, denied public egress, routes verified continuously.
Beside a productivity assistant
Where a client is rolling out a general productivity assistant, Shelby takes the workloads whose data should not go to one. The two run side by side.
Evidence your client can read
The Shelby Register records, per model version, where it may be processed and under which retention regime.
Partner pilot format Bring three client opportunities. Each is scoped as its own 30-day pilot, acceptance criteria fixed before kickoff.
Who it is for
Organisations that hold sensitive data and answer to a regulator or a client.
Insurers, brokers and MGAs
Quoting, bordereaux and management reporting, DORA artefact production, invoice processing and client-money reconciliation.
Mortgage and credit intermediaries
Application handling and document checks, kept to preparation, never scoring.
Contractors and MEP firms
Tender clarification and QA from drawing sets, and progress reporting.
Public bodies
Casework preparation, records handling and reporting, kept on premises with a tamper-evident audit trail.
Enterprises
The same blueprints at department scale. Runs beside a general productivity assistant and takes the work that assistant should not see.
Research · Amata Consulting imprint
The regulation, read closely.
The work behind Shelby’s controls, published for compliance teams, policymakers and peers.
- Policy paperReconciling Ireland’s AI Ambition
- Market briefingThe Governance-Confidence Gap — response to the DETE sectoral consultation on enterprise AI adoption
- Market briefingArticle 50: The Deadline That Survived
- Advisory familyDORA: overview, AI on the Register of Information, RoI and exit-plan build, vendor pack
- Case studiesDORA worked examples: vendor pack, greenfield build, remediation
Team
Four people, from the switchroom to the model.
Silviu Preoteasa
Strategy and product
Product strategy, financial modelling and architecture decisions.
Daryll Collins
Engineering and delivery
MEP engineering, client delivery and field operations.
Jochen Cuntz
Systems integration
APIs, data pipelines and agent orchestration. Wires blueprints into your systems.
Marian Cursaru
Security architecture
Infrastructure hardening, network operations and the security posture regulated buyers need.
Scope a pilot
Tell us which workflow you want off your team’s desk.
We reply with a draft scope: the blueprint, the data boundary and the acceptance tests. Or write to hello@aifactor.ie.