On-premise AI
Your data cannot leave the company. You can still use AI.
Open-source models running in your own server room, connected to your ERP and company systems, available to staff in the tools they already use. We do not bill by the day: we write down the result first, and you pay when it runs.
- 7 yrs
- in business
- 252+
- projects
- 32+
- industries
- NT$20M+
- delivered in one year
You may be stuck here
- Staff already paste client data into ChatGPT, and nobody knows where it went.
- You want an AI knowledge base, but contracts or audits forbid the cloud.
- You bought an AI server and it never got connected to anything.
What you get
AI in your server room, connected to your systems
- 01Works offline
Runs in your server room, even with the internet cut off.
- 02Answers with sources
A company knowledge base that cites where each answer came from.
- 03No new tool to learn
Connected to your ERP, company systems and chat.
- 04Permissions and logs
Access by department; every question is logged.
Up front
- If off-the-shelf tools can do it, we will tell you not to build.Custom systems are not cheap. If existing software is enough, a spreadsheet tweak will do, or the process itself needs fixing first, we say so on the first call.
- You get the source code and the database schema.You paid for the system, so it is yours. Annual maintenance with us is a separate agreement, and it is your call.
- Tweaks within the original design are free during maintenance.Things nobody thought of in the interviews, or that only show up once you are using it: if they fall within the delivered feature list, changing fields, screens or reports costs nothing extra. Critical issues are fixed within 24 hours of the report, high-risk within 72. The system grows with your company.
- Every system ships with in-app guides and a user manual.Each system includes step-by-step in-app guidance and a full manual, and we walk your team through it in person at launch. When someone leaves or a new hire starts, they open the system and follow the guide.
Our work
Medical and client data that cannot go to the cloud
Hospital records, nursing systems, offline AI support, a private model for a consultancy.
How to start
Four steps, and you can stop at any of them
We do not bill by the day. We write down the result first, and you pay when it runs.
- STEP 01Free 30-minute call
Talk with our founder about whether your data and needs suit on-premise.
- STEP 02Free AI trial
Send 20 de-identified real questions; we show you the answers within a week.
- STEP 03Proof of concept (first contract milestone)
Test the model, data, hardware, integration and value; stop if it misses the bar and pay only the first milestone.
- STEP 04Launch and acceptance
Pricing
- Implementation
- Quoted by scale
- Hardware
- At costNo markup: we write the spec, you compare prices yourself.
- Maintenance
- From NT$50,000/monthA company running on-premise AI needs an IT team to look after it. We act as that team: model updates, knowledge base upkeep, security patches, incident handling, monthly reports.
Who it is not for
If your data is not sensitive and your volume is low, cloud AI is cheaper and smarter, and we will tell you so.
Further reading: on-premise AI series (Chinese)
- How to choose an on-premise AI model: rule out what you can't use, then test with your own questions
- What hardware does on-prem AI need? Tiers and prices from RTX 5090 and Mac to H200
- How fast is on-premise AI? One user vs. dozens at once
- What on-premise AI can't do: 7 limits owners should know before implementation
- Open-source models or Claude, ChatGPT, Gemini, Grok? First ask whether your data can leave the company
- Should you fine-tune an AI model? Most companies only need a knowledge base first
- Running company AI on a Mac: Mac Studio pros and cons for local models
- Thinking of building an AI data center to sell compute? Power, money, and Taiwan's grid limits
- How an 8B model beat top frontier models at its own job: breaking down NVIDIA's fine-tuning workshop
- 9 checks before you fine-tune an AI model: don't spend weeks training one you can't use
FAQ
Five common questions
Is on-premise AI dumber than ChatGPT?
On complex reasoning it may trail the best cloud models; for searching company data, organising, summarising and drafting it is enough — the trial tests it on your own questions.
Do we have to buy hardware?
The trial and proof of concept run on rented machines; buy only once it is proven. You buy the hardware yourself, and we take no margin.
Is on-premise automatically secure?
No. It needs permissions, logs and regular updates, all included in maintenance.
What happens when the model gets outdated?
Open-source models update every few months; during maintenance we test new versions on your questions and switch only if they are better.
Can we train our own model?
Most companies do not need to. Organising your data into a knowledge base usually works better and costs far less.
Next step