The short answer
We deployed private AI for a real estate business in the UAE on servers in the client's own premises. It serves up to 750 users, processes transaction records and images for about 5,000 clients, and took eight to ten weeks from the first design to launch. Most of that time went on the architecture and the security, before any model was switched on.
- Up to 750 users, on the client's own servers
- Transactions and images for about 5,000 clients
- Eight to ten weeks from design to launch
- Architecture and security designed before deployment
What the client was worried about
Before the first call, the worry was the hardware. Like most businesses, they assumed private AI needs the largest, most powerful servers on the market.
It does not. A GPU is still required, but many current models adapt to far more modest hardware than a business expects. The servers were chosen to the client's requirement, to fit their users and their data.
What was built
We built a model on the client's own private servers, on premise, and trained it on their own records. No part of it is public. It exists only for internal use.
It processes property transaction records and images for about 5,000 of their clients. The business uses the results for in-depth analytics and reporting, so that it can see where to improve revenue and productivity.
The eight stages
This is how the eight to ten weeks were spent.
- Draft the solution
- Complete the solution architecture: where each service sits and how data flows between them
- Present the architecture to the stakeholders and agree it
- Set up the servers
- Deploy each service
- Put the security in place
- Test run on a small data set
- Launch
The one thing we would tell you not to do
Do not take a server and deploy your own private AI before the architecture and the security are designed.
Decide first where each service will sit, how data flows in every situation, and which firewalls protect it. That is a solutions architect's job, and it comes before the hardware, not after.
If you are considering the same
Start with your users and your data, not with a server specification. Tell us how many people would use the system and what documents it would read, and we will set out the architecture before anything is bought or installed.