Deployment options
| Option | What it means | Good fit when |
|---|---|---|
| Managed model in your cloud tenancy | Models served through your cloud provider's AI service in your own account and region (for example Azure OpenAI, Amazon Bedrock or Google Vertex AI) | You need strong data controls and regional residency but want managed infrastructure |
| Self-hosted open-weight models | Models such as Llama, Mistral or Qwen served on GPUs in your VPC with an inference server like vLLM | Data must not leave your network, or you need full control over model versions and cost |
| On-premises or air-gapped | Open-weight models on hardware in your data centre with no external connectivity | Regulatory, defence or sovereignty requirements rule out cloud processing |
| Hybrid | Sensitive workloads on private models; lower-risk workloads on managed services | Different use cases have different risk and cost profiles |
What we deliver
- 01
Use-case and model selection
We evaluate candidate models on your own tasks and data, weighing quality, latency, cost and licence terms, rather than choosing on public benchmarks.
- 02
Secure hosting
Infrastructure as code for the serving stack, private networking, encryption, secrets management and autoscaling, deployed into your accounts.
- 03
Retrieval over your data
Pipelines that index documents and records with the same access permissions as the source systems, so users only get answers from data they are allowed to see.
- 04
Access control and gateway
A model gateway with single sign-on, role-based access, rate limits, usage tracking and prompt and response logging.
- 05
Evaluation and guardrails
Task-specific evaluation sets, output filtering and policy checks, re-run before every model or prompt change.
- 06
Monitoring and handover
Dashboards for quality, latency, cost and usage, runbooks, and training so your team can operate and extend the system.
Security and compliance
- We never use your data to train models for anyone else, and we configure managed model services with the provider's data-retention and training opt-outs. Your data and fine-tuned weights remain your property.
- Data residency by region, with processing kept inside your chosen boundary.
- Permission-aware retrieval that mirrors source-system access.
- Full audit logging of prompts, retrieved context and responses, with retention you control.
- Security controls and data-flow documentation you can map to your own compliance frameworks, handed over with the system.
How an engagement runs
We begin with the use cases and constraints, run a model bake-off on your data, and deploy the platform and first application together. The standard is a working prototype on your data within 3 to 5 weeks and production within about 60 days. You own the code, infrastructure definitions and any fine-tuned models; work is scoped as fixed-price milestones.
Frequently asked questions
Are open-weight models good enough compared with hosted frontier models?
For many enterprise tasks, such as classification, extraction, summarisation and retrieval-based Q&A, well-chosen open-weight models perform well, especially with retrieval and task-specific evaluation. For the hardest reasoning tasks, hosted frontier models may still lead. We test both on your own tasks before recommending.
Can we keep data inside our own cloud account?
Yes. We deploy into your cloud accounts, either using the provider's managed AI service in your tenancy and region or by self-hosting models on GPUs in your VPC.
What does a private LLM cost to run?
It depends on model size, traffic and latency requirements. Self-hosting has a fixed GPU cost that suits steady, high-volume use; managed services charge per token and suit variable use. We model both on your expected usage before you commit.
Do we need to fine-tune a model?
Usually not at first. Retrieval over your data plus good prompts and evaluation covers most use cases. Fine-tuning is worth it when you need a consistent format or style, domain vocabulary, or a smaller, cheaper model to match a larger one on a narrow task.
Who operates the system after launch?
Your team, with runbooks, dashboards and training from us. We can also provide ongoing support and model updates if you prefer.