Private AI
We set up local AI for clinics, law offices and financial firms that cannot send client records to services like ChatGPT. The model runs on hardware you control, answers from your own documents and logs every question.
If a hosted service would do the job, we say so on the first call.
- Runs on
- Your servers or a private cloud you own
- Models
- Llama, Mistral or Qwen
- Your data
- Stays inside your network
- Access
- By staff role, every question logged
- Price
- Quoted per project
- The setup
- Yours, with operating instructions
What we set up
The model, the index of your documents and every question stay inside your network. Nothing is sent to a hosted AI service.
-
A model on your hardware
Open-weight models such as Llama, Mistral and Qwen can be downloaded and run on your own machines. We install one sized to your hardware and the work, on servers you own or in a private cloud tenancy in your name.
-
Answers from your own documents
The model answers from your files: policies, records, contracts and your knowledge base. Each answer names the document it came from, so staff can check the source before they rely on it.
-
Connections to your tools
It works inside the systems your team already has, such as internal search, a support inbox, an intake form or a line-of-business app. We build the connection against what you run today, so nothing has to be replaced.
-
Access control and audit logs
Access is set by staff role. Each question is logged with who asked it, when, the documents it drew on and the answer given. When an auditor asks who saw what and when, you can export the record.
Who it is for
Organizations where an outside AI service seeing client data would be a regulatory or contractual problem.
-
Clinics
Clinics and health teams in Ontario that hold patient records under PHIPA and cannot send them to OpenAI or Anthropic. Drafting, summaries and search over charts run on infrastructure the clinic controls. See software for clinics.
-
Law offices
Firms and in-house legal teams bound by privilege and confidentiality, who need search and drafting over matter files without those files reaching an outside model or its training data.
-
Financial firms
Teams holding client financial data under contracts or data-residency rules that forbid sending it to a hosted AI service. The model and the data stay on infrastructure the firm owns.
When it makes sense
Most businesses do not need this.
If your data is not regulated, a hosted service such as OpenAI or Anthropic is cheaper, quicker to set up and usually a stronger model than anything you would run yourself.
Worth it when
- A privacy law, a regulator or a contract forbids sending the data to a third party.
- Rules say where the data must be stored and processed.
- You have a specific reason the data must never leave your network.
- Your team will use it at a steady volume.
A hosted service is better when
- Your data is not regulated, and an AI provider's data terms are acceptable to you.
- Use is light, or comes in occasional bursts.
- You need the strongest open-ended reasoning available.
Cost
There is no flat price. Before you commit, we write up what running your own model would cost for your usage, set against a hosted service.
- Setup
- Quoted per project, in writing
- Running costs
- Hardware, power and upkeep, instead of paying per use
- What sets the cost
- The hardware, the size of the model and how many people use it
- After launch
- A monthly retainer, or your team runs it
- Ownership
- Yours, with the configuration and operating instructions
Further reading
-
When local AI deployment is the right answer
The situations where sending data to a cloud AI provider is not an option, and when running your own model is worth the extra work.
-
The break-even point for running LLMs yourself instead of cloud APIs
Rough numbers on the point where running a model yourself starts to cost less than paying a provider per use.
-
The total cost of self-hosting an LLM in 2026, from hardware to engineering time
Every cost of running your own model, from hardware and power to upkeep, set against a per-use bill.
-
PHIPA-compliant AI for Ontario clinics: a build checklist
What an Ontario clinic has to get right before using AI with patient information.
Questions
Does our data ever leave our infrastructure?
No. The model, the index of your documents and the processing all run on hardware you control, either on your premises or in a private cloud tenancy you own. Prompts, retrieved documents and the index stay inside your network. When the system answers a question, it does not call OpenAI, Anthropic or any other hosted AI service.
Which models do you use, and are they good enough?
Open-weight models: Llama, Mistral and Qwen, sized to your hardware and the task. For answering questions from your own documents, a mid-size open model usually does the job. If your use case needs more than that, we will tell you before you buy any hardware.
What does it cost to run?
It depends on the hardware, the size of the model and how many people use it. Running your own model replaces a per-use bill with fixed costs for graphics processors, power and upkeep. That makes sense at steady volume, or when the rules you work under exclude cloud services. It makes less sense for light or irregular use. We write up the trade-off for your case before you commit.
How does this fit with PHIPA and similar privacy rules?
Keeping patient or client data on infrastructure you control, with access controls and audit logs, lines up with how regulated organizations are expected to handle sensitive records under PHIPA and similar rules. We build the technical controls: the data stays inside your network, access is set by role, and queries are logged. We are developers. We are not your lawyers or your privacy officer, and the final compliance sign-off is theirs to give.
Can you maintain it after it goes live?
Yes, on a monthly retainer: model updates, security patches on the server, refreshing the document index as your files change, and tuning as use grows. You can also take it in-house. You own the deployment, the configuration and the written operating instructions, and nothing depends on our infrastructure.
Start a project
Tell us what you want the AI to do.
Send a few lines about your organization, the documents involved and the rules you work under. We reply within one business day, and if a hosted service would do the job, we say so on the first call.
Or call 647-252-5352, Monday to Friday, 9 am to 6 pm Toronto time.
