Medical records and client data can't go to the cloud. You can still use AI — four on-premise AI projects Brightstream has built

A clinic's medical records and a consulting firm's client data can't be sent to cloud AI. We've done four projects like this: a medical records system, a nursing system, an offline AI customer service, and an on-premise LLM for a consulting firm. The data never leaves the company.

Client
Healthcare Providers and Consulting Firms
Industry
AI Native
Service
Custom AI systems
Medical records and client data can't go to the cloud. You can still use AI — four on-premise AI projects Brightstream has built
Results

What we delivered.

Data
Never leaves the company
Projects
Medical records system, nursing system, offline AI customer service, on-premise model for a consulting firm
Connectivity
Runs with the outside internet cut off
Governance
Tiered permissions + query logs
Story

01The companies that most want AI are often the ones that can least use cloud AI

Healthcare providers hold medical records. Consulting firms hold their clients' financials and contracts. This is exactly the data that most needs AI to organise and search it. It is also exactly the data that can least be sent to an outside cloud.

None of the clients in these four projects agreed to be named, so we only describe the industry and the approach.

02The four projects

1. Medical records system for a healthcare provider

Medical records are the most sensitive personal data there is. In the medical records system we built, "who can see what" came first from day one: permissions tiered by role, a log entry for every single access, and data kept in an environment the provider controls itself.

2. Nursing system

Clinical workflows are very different from a normal company's: shift handovers, scheduling, care records. Time pressure is high and mistakes are not allowed. This project gave us a clear grasp of clinical workflows and data rules, and it became the foundation for our later medical AI work.

3. Offline AI customer service

The client's environment can't connect to the outside internet, so the AI customer service had to run on its own internally. We installed an open-source model and the company's knowledge base on an internal server. Customer Q&A keeps working with the outside internet cut off. Answers cite their sources, so staff can go back and check.

4. On-premise LLM build for a consulting firm

The consulting firm holds clients' accounts, contracts and reports, and the contracts explicitly forbid that data from leaving. We built an open-source LLM in their own server room and connected it to internal documents. Consultants use AI to look things up, summarise, and draft reports. The data never leaves the company.

03What all four projects had in common

  1. First confirm which data can't leave, then pick the model and hardware. Not every company needs on-premise. If the data isn't sensitive and usage is low, we'll tell you to just use the cloud.
  2. Test with your own questions first: during model selection we test with the client's real questions. Traditional Chinese and industry terminology both have to pass.
  3. Plug into existing systems: AI isn't a separate web page. It sits inside the systems staff already use every day.
  4. Permissions and logs: what each department can ask and see is configured by department. Every query is logged and can be audited.
  5. Works with the network down: the model, knowledge base and inference service are all internal. Nothing depends on outside services.

04Further reading

How to choose an on-premise model, what hardware to buy, and what the limits are — we've written these up as a series: How to choose an on-premise AI model, What hardware on-premise AI needs, What on-premise AI can't do, Open-source models or cloud LLMs.

05My take

The point of on-premise AI isn't "getting the model installed". Installing a model takes a day. The point is connecting it to company data, plugging it into the workflow, and getting permissions and logs right, so that companies whose data can't leave can actually use AI every day too.

Meet Johnny Yang →

Your projectTalk for 30 minutes,then decide.
Free 30-minute diagnostic · pick a slot online · with Johnny himselfIf, after the call, we do not think we are the right fit, we will tell you honestly.Or ask on LINE →