AI development company for Cardiff fintech, insurance, media, gaming, and Welsh public sector.

No — Aiinfox doesn't keep a Cardiff office; delivery runs out of our Mohali, India HQ and a Frisco, TX office, with four to five hours of native daily overlap against UK business hours built into the Mohali schedule. When a milestone review, security walk-through, or kickoff genuinely calls for a room together — at Newport Road, Cardiff Bay, the BBC Wales building on Central Square, Cardiff University, or Welsh Government's Crown Buildings — we fly in on a planned cadence rather than carrying a sub-scale local desk. The senior engineer who runs your kickoff call is the same one shipping your code at launch.

Honest answer: we deliver bilingual customer-facing AI with a fluent Welsh-speaking reviewer on the output, not a fluent Welsh-speaking engineer on the development team. For voice agents, we use Deepgram for Welsh STT where available and ElevenLabs Welsh voices for TTS, with validation passes against native speakers before any production release. For chat agents and RAG, the model handles Welsh and English generation; we run a Welsh-language reviewer over a representative sample of customer-facing output during week one of the build, then again before launch, and continuously on a sampled basis post-launch. For document intelligence over bilingual corpora, we use models with strong Welsh coverage and validate extraction accuracy against a held-out Welsh-language test set. If your engagement requires a fluent Welsh-speaking engineer embedded in the build rather than a reviewer on the output, we will be straight on the first call and recommend a Cardiff or Aberystwyth-based bilingual delivery partner.

Yes. Admiral, Principality, and the broader Cardiff financial-services cluster operate under FCA Senior Managers and Certification Regime, and we mirror the SM&CR audit-trail patterns so the senior manager accountable for the system has the evidence needed for a Section 166 review or a routine supervisory examination. Every model and tool call is audit-logged with input, output, prompt version, and operator identity. For motor-insurance and claims-handling workflows specifically — Admiral's anchor business — we have shipped outbound voice, claims-triage document intelligence, and fraud signal extraction at insurance-cluster scale. Humans approve everything that touches a regulated outcome.

Yes. Our engagement defaults are aligned with UK GDPR and the Data Protection Act 2018. DPAs (UK GDPR controller-to-processor or processor-to-processor as appropriate) are signed before any personal data is shared. For engagements processing personal data at scale or involving special category data, we run a Data Protection Impact Assessment (DPIA) aligned with the ICO's published guidance. The UK International Data Transfer Agreement (IDTA) or the EU SCCs with the UK addendum cover any necessary international transfer of personal data. Audit logs on every model and tool call are exportable for ICO inspection.

Yes. NHS Wales operates a distinct information governance pattern from NHS England — the NHS Wales IG framework, the Caldicott Guardian principles as applied within NHS Wales, and the data-sharing protocols inside the Welsh public sector. For Cardiff University Hospital-adjacent and broader NHS Wales-adjacent healthtech, we map the data flow against the Caldicott seven, document the lawful basis under UK GDPR for any special category processing, and ensure audit logs on every model and tool call are exportable for IG officer review at the contracting Health Board. Deployments default to AWS London or Azure UK South inside customer-controlled UK clouds.

You choose the region. AWS London (eu-west-2), Azure UK South, and GCP europe-west2 are our standard defaults, and we'll run the full build inside your own UK cloud account if your DPO wants to rule out cross-region replication entirely. For inference we route Claude and GPT-4o to UK or EU endpoints where the provider offers one — Azure OpenAI UK South is the usual path for GPT-4o — and for bilingual Welsh + English corpora, or anywhere third-party inference is off the table, Llama 3 runs self-hosted on vLLM inside your VPC. The DPA sets out exactly which dataset goes where.

Most v1 engagements at Aiinfox land between GBP 20,000 and GBP 100,000 fixed-price for a focused build — an AI agent, a RAG system, a voice pipeline, or a bespoke ML model. Larger multi-quarter engagements with fine-tuning, custom evals, FCA-aware insurance work, or NHS Wales IG-aligned compliance work typically reach GBP 150,000 to GBP 280,000. Bilingual Welsh + English delivery adds roughly 10 to 20 percent to the scope depending on the depth of native-reviewer engagement required. Pilots are usually GBP 8,000 to GBP 20,000 with acceptance criteria written into scope. Cardiff clients invoice in GBP via bank transfer.

Yes — and the audit usually surfaces the same gap: the Welsh-language requirement got treated as a translation checkbox rather than a build requirement, so bilingual output was never actually validated by a native reviewer before launch. We start by reading the code, the data pipelines, the eval results if any exist, the prompts, and — where Welsh output is in scope — whether a native Welsh reviewer ever signed off on production content. The first shipped change is often that validation pass itself, which proves we understand the system faster than the previous vendor did. From there it's a straight call: incremental stabilisation, a parallel rebuild, or a clean shutdown. Most Cardiff takeovers we have seen stalled less on the AI engineering and more on a London-based vendor underestimating what Admiral-adjacent insurance rigour or BBC Wales-adjacent bilingual delivery actually required. We will tell you which applies to your build on the first call.