Private AI & On-premise LLMsAI that runs where your data lives.
Open-source language models such as Llama, Mistral and Qwen deployed on your own servers or private cloud, with assistants, search and document AI built on top. For organisations whose data cannot be sent to public AI services.
- Your servers or private cloud
- Open-source models
- No data sent to AI vendors
- Access control & logging
- Contract summary generated · on-premise model
- Access denied · user lacks HR permission
- Model update tested against evaluation set
- GPU usage at 61% · within plan
Some data should never leave the building.
Public AI services are convenient, but many regulators, clients and contracts do not allow sensitive data to be sent to them. Private AI gives your team the same tools inside your own boundary.
Stays on your servers or private cloud
Sent to a vendor, often abroad
Fits data residency and sector rules
Hard to justify to regulators
Your identity system and permissions
Shared accounts, little oversight
Predictable infrastructure cost at volume
Per-user or per-token fees that grow
Models tuned to your documents and language
One model for everyone
Every prompt and answer logged for review
Limited usage logs
Designed so prompts and documents never leave your network.
We benchmark models on your tasks and size the hardware to match.
Tested by our in-house AI security team.
ISO/IEC 27001:2022 certified. Your code and data are handled under audited controls.
Who needs private AI
Private AI suits organisations that want the productivity of AI assistants but cannot send their data to public services. Common examples are hospitals and clinics, banks and insurers, law firms, government bodies, defence suppliers and companies bound by client confidentiality or data residency rules.
What these buyers need
- Certainty about where data goes, written into the architecture rather than a vendor’s policy.
- Answers that respect existing permissions, so AI does not become a way around access control.
- Logs that compliance and security teams can review.
- Predictable running costs as usage grows.
- Honest advice on which tasks open-source models handle well today.
Deployment options
- On-premise: models run on GPU servers in your own data centre.
- Private cloud: models run in your own AWS, Azure, Google Cloud or regional cloud account, in the region you choose.
- Hybrid: sensitive tasks run privately, while low-risk tasks may use hosted models under enterprise terms.
Questions to ask any provider
- Which models will you test, and on which of our tasks?
- What hardware or cloud cost should we expect at our usage?
- How are prompts, answers and documents logged and retained?
- How will models be updated without breaking what works?
Sigosoft works under ISO/IEC 27001:2022 controls, and our in-house security team tests AI systems for prompt injection and data leakage before they go live.
What we deploy privately.
Private ChatGPT-style assistant
A secure chat tool for staff, on your infrastructure.
Document Q&A
Ask questions across contracts, policies and records.
Document processing
Extract data from sensitive files without sending them out.
Coding assistants
Help developers without exposing your source code.
Fine-tuned models
Models adapted to your terminology and tasks.
GPU infrastructure
Sizing, setup and monitoring of on-premise or private-cloud GPUs.
Benchmark first. Then buy hardware.
- 1Weeks 1–2
Assess
Use cases, data rules and a benchmark of candidate models on your tasks.
- 2Weeks 3–4
Size
Hardware or private-cloud sizing with running-cost estimates.
- 3Build
Deploy
Models, access control, logging and the first application.
- 4Ongoing
Operate
Model updates, evaluation and monitoring under SLA.
AI we have put into production.
“Third agency we tried. First one that actually delivered.”Khalid Al-Rashid · Head of Operations, StratEdge GroupView all projects
For many business tasks, such as summarising, answering from documents and extracting data, current open-source models perform well. We benchmark candidate models on your real tasks before you invest, and say plainly where a hosted model would do better.
It depends on the model size, the number of users and the response speed you need. Some use cases run on a single GPU server; others need more. We size it after benchmarking, and can also deploy in a private cloud tenancy instead of buying hardware.
Yes. We deploy to AWS, Azure, Google Cloud or regional cloud providers inside your account and region, or on servers in your own data centre.
The assistant connects to your identity system and respects document permissions, so users only get answers from content they are allowed to see. All prompts and answers are logged.
We do, under a support agreement. New model versions are tested against your evaluation set before they replace the current one.
Give your team AI without the data risk.
Tell us what your team wants AI for and which data it would touch. We reply with a deployment option and cost range.

