Private AI
AI that runs where your data already lives.
Public AI services are useful. They are also someone else’s computer. Every document you upload and every question you ask leaves your organisation — and for client files, staff records or anything under an NDA, that’s a question you’d rather not have to answer.
What it is
A complete AI stack, inside your walls.
Kynthos deploys a full AI system on infrastructure you control — a server in your office, or your own cloud account — running open-weight models. Your apps talk to it through an OpenAI-compatible API, so most existing tools and code work with little or no change.
Everything is pinned and written down: engine builds, model versions, configuration. What you test is what you run.
Your infrastructure — office server or your own cloud account
- AskYour apps & staffChat, documents, internal tools
- Check & routeGatewayOne OpenAI-compatible API · keys · limitsLiteLLM
- AnswerOpen-weight modelsChosen per task: chat, search, extraction, visionvLLM on NVIDIA · llama.cpp
Watching every request
- TracingOpenTelemetry → Langfuse
- Metrics & alertsPrometheus · Grafana
- EvaluationsInspect AI
Nothing is sent to outside AI providers.
What’s in the box
Four parts, all yours.
- Models
- Open-weight models chosen for each job: chat, document search, extraction, and reading scans and drawings.
- Gateway
- One OpenAI-compatible endpoint with access keys, usage limits and logs that belong to you. Swap models without touching your apps.
- Tracing and metrics
- See every request, how long it took and how hard the hardware worked. Alerts when something drifts.
- Evaluations
- Test sets built from your own work, so you know whether the AI is good enough before your staff rely on it.
Typical uses
Where it earns its keep.
Examples of the kind of work it suits — not client case studies.
- 01
Answering questions over internal documents, policies and manuals
- 02
Drafting replies, quotes and reports from your own templates
- 03
Pulling data out of invoices, forms, scans and PDFs into your systems
- 04
An internal assistant for staff that never touches the public internet
- 05
Agentic workflows that carry a task through several steps, with a person approving the result
- 06
Fine-tuning open models on your own examples, so they learn your terms, formats and the rules of thumb your team works by
How we get there
Assess, pilot, run.
- 01
Assess
Which tasks, which data, and what hardware or cloud budget makes sense.
- 02
Pilot
One workflow, end to end, as a fixed-scope proof of concept — with evaluations that show whether it works.
- 03
Run
Hand it over to your team with documentation, or Kynthos keeps it running, updated and monitored for a monthly fee.
Questions
Frequently asked
Does any of my data leave our infrastructure?
Not for the AI itself. Models, prompts, documents and logs all stay inside your environment, and there is no fallback to outside AI services. Software and model updates are brought in; nothing is sent out.
Can it run in our cloud instead of our office?
Yes. It runs either on a server at your premises or in your own cloud account (for example in a Singapore region), on GPUs that only you control.
Does this make us PDPA-compliant?
It removes a whole class of risk — your data isn’t sent to an overseas AI provider — but the PDPA covers how you handle personal data overall. You get documentation of how the system handles data, so you can fit it into your own policies.
Are open models as good as ChatGPT or Claude?
For many business tasks — searching documents, drafting, pulling data out of forms — they are good enough, and the evaluations show it on your own work. They are not the largest models in the world; where a hosted model is the better choice, you’ll hear that up front.
What if the server goes down?
A single server is a single point of failure. Sizing, redundancy and the uptime it can realistically promise are agreed before you commit. In your own cloud account, capacity can be added when it’s needed.
What hardware do we need?
It depends on the models and how many people use them. Sizing is part of the assessment, before you buy or rent anything.
Can we change models later?
Yes. Everything talks to the gateway through one standard API, so models can be swapped or upgraded without changing your apps.
Get in touch
Talk about a private AI pilot.
A few lines about your business and the problem are enough. You’ll get a reply within one business day.