Skip to content
CBN data localisation deadline, 1 Jan 2027: 99d 00h 31m left.Talk to us →
NuxFamily

Generative AI on your most sensitive data, on GPUs in Nigeria, with no data leaving the building.

Run AI on your most sensitive data without your data ever leaving the building

On-premise and air-gapped AI for banks and regulated institutions: GPU infrastructure, open-weight models, private RAG and MLOps, deployed in your Nigerian data centre and supported as part of the same private cloud.

Book a Discovery SessionSee the CBN compliance path

The challenge

The challenge

Banks want the productivity of large language models on their own documents, procedures, contracts and transaction history. Almost all of that material is either payment data covered by the CBN localisation directive or personal data under the NDPA 2023. Sending it to a foreign AI API is not an option. This is context, not legal advice.

The journey with NuxFamily

  1. 01Assess
  2. 02Design
  3. 03Build
  4. 04Migrate
  5. 05Operate
  6. 06Evolve

Every family is delivered through the same six-stage journey, with official 24×7 support and knowledge transfer built in.

Our approach

How we deliver it

  1. 01

    Assess the use cases

    Candidate use cases ranked by value, data sensitivity and connectivity constraint. Each one gets a deployment class: on-premise connected, on-premise restricted or fully air-gapped.

  2. 02

    Size the GPU platform

    Accelerators, memory and storage sized against model choice and expected concurrency, not against a vendor's reference sheet. Shared scheduling on Kubernetes so no GPU sits idle.

  3. 03

    Build the inference layer

    Open-weight models served through an inference server with quotas, observability and an API compatible with what application teams already use.

  4. 04

    Ground it in your data

    Private RAG over documents, policies and structured data: ingestion, embeddings, vector search in OpenSearch and access control that respects existing permissions.

  5. 05

    Lock down the air gap

    For disconnected workloads: internal registry, signed model artefacts, supply-chain verification and a documented chain of custody for every model update.

  6. 06

    Operate and transfer

    MLOps pipelines for evaluation, versioning and rollback. Official support on the platform, and your data science team trained to run it.

Outcomes

Outcomes

0

bytes sent to external AI APIs

Prompts, documents and embeddings are processed and stored on your own infrastructure.

100 %

air-gap capable

Models, dependencies and updates delivered through a controlled channel with chain of custody.

1

platform for AI and data

The AI layer runs on the same Kubernetes, storage and data services as the rest of the private cloud.

20+

years in regulated infrastructure

The same discipline used for European banks such as Santander, ING and Bankinter, applied to AI.

Use cases

Use cases

Use case 01

Private assistant over policies and regulation

A bank wants an internal assistant that answers staff questions about procedures, CBN circulars and product terms, citing the source. None of the documents may leave the bank's network.

Technologies

  • Open-weight LLM
  • GPU inference server
  • OpenSearch
  • Apache Kafka
  • Kubernetes
  • MinIO
  • OpenMetadata

Expected outcome

Staff get sourced answers in seconds; the bank keeps every document, prompt and answer inside its own perimeter.

Metric: Answers grounded in cited sources, 0 external calls

Private assistant over policies and regulationA bank wants an internal assistant that answers staff questions about procedures, CBN circulars and product terms, citing the source. None of the documents may leave the bank's network. Staff get sourced answers in seconds; the bank keeps every document, prompt and answer inside its own perimeter.1Ingest documentsApache Kafka2Chunk and embedEmbedding model on GPU3Index for retrievalOpenSearch4Generate with citationsInference server5Evaluate and monitorMLOps pipeline6Audit every answerOpenSearch
  1. 1Ingest documents. Policies, circulars and manuals pulled from document stores with permissions preserved per document.
  2. 2Chunk and embed. Text split, embedded with an open-weight model and stored with metadata for filtering.
  3. 3Index for retrieval. Vectors and keywords indexed together; hybrid search returns passages the user is allowed to see.
  4. 4Generate with citations. An open-weight LLM answers from retrieved passages only and links each claim to its source.
  5. 5Evaluate and monitor. A test set of real staff questions scored on every model or prompt change before release.
  6. 6Audit every answer. Prompts, retrieved passages and responses logged in-country with retention for review.

Use case 02

Air-gapped anomaly scoring in the payment flow

Built on these families

Built on these families

FAQ

FAQ

Open-weight models whose licences permit commercial use, selected for your language, task and hardware. We evaluate candidates on your own test set before recommending one.

Over two decades

Built by the team behind the platforms of Santander, ING, Bankinter, Mapfre and Inditex

More than twenty years designing, building and operating private clouds for institutions that cannot afford to fail, and a delivery model where we stay with you from assessment to operation.

See our track record

20+

Years building private clouds

40+

Private clouds delivered

Start with a Discovery Session

Sixty minutes with an architect who has delivered this before. No slideware.

No mailing lists, no automated follow-ups. We reply personally within one working day.