Private AI for Organisations, Inside Your Building

1. 1

   Message us on WhatsApp

   Tell us which work eats staff time and what languages your documents are in.

2. 2

   Choose real samples

   Pick 20–50 documents or recordings per task, hard ones included. Pass criteria and handling rules are agreed first.

3. 3

   Review the pilot results

   Results next to the originals, mistakes included. We say plainly which tasks are ready and which are not.

4. 4

   Size the server

   Based on the tasks that passed, the number of users and the location. The quote names brand and model.

5. 5

   Install on site

   Server, UPS, user accounts, permissions, usage logs and backup.

6. 6

   Acceptance test

   Run the test set agreed in writing. Go live only when it passes.

7. 7

   Monthly care

   Updates, test re-runs, new documents added and help on WhatsApp.

Private AI usually runs on one of three sizes of machine. We choose the size after the pilot, based on your work and number of users — no fixed brand.

Photo: Daniel Lu, CC BY-SA 4.0 · cropped and resized

Desktop AI box

Looks like

Smaller than a shoebox; sits on a desk or shelf in the office and runs quietly.

Good for

One team or department, pilots, document search and transcription.

You need

A normal power socket, the office network and a small UPS. No server room needed.

Example models

For example NVIDIA DGX Spark and machines from other makers built on the same chip.

Photo: Daniel Lu, CC BY-SA 4.0 · cropped and resized

Several boxes linked together

Looks like

Two or more desktop boxes on one shelf, joined through the high-speed network ports on the back (pictured).

Good for

Several departments at once, models too large for one box, or one job per box.

You need

A high-speed network switch, a larger UPS and a well-ventilated spot.

Example models

For example several DGX Spark units linked together.

GPU server in a rack

Looks like

A flat server mounted in a rack with AI processing cards inside; its fans are louder.

Good for

The whole organisation, many users at once, work that runs day and night.

You need

A server room or rack, server-grade power and UPS, and cooling.

Example models

For example servers with NVIDIA data-centre cards from mainstream server makers.

The first two images are photos of the real machine (credit under each image); the third is an AI-generated illustration. NVIDIA and DGX are trademarks of NVIDIA Corporation. We are not an official NVIDIA partner.

• AI can be confidently wrong and invent numbers or names. We do not claim the output is always right, which is why a person checks it.
• Lao in AI is still weaker than Thai and English for many tasks. Handwriting, text mixing three languages and poor scans are the hardest cases.
• Transcription struggles with overlapping voices, noisy rooms and strong regional accents. A good microphone helps more than software.
• Answers are only as good as the documents. If folders hold conflicting versions or blurry scans, the AI cannot do better than that.
• AI does not replace staff. It drafts and searches, but decisions and responsibility stay with people.
• AI images and video are always labelled as AI-generated and are never presented as the real product, place or person.
• Some tasks are not ready for AI yet. If the pilot shows that, we tell you plainly before you buy any hardware.

• Pilot and acceptance test results, side by side with the originals
• The agreed test set and pass criteria, for re-running after every update
• User guide in Lao
• Administrator guide
• List of user accounts and the rights of each group
• Backup and restore procedure, with the result of a real restore test
• Equipment list with brand, model and warranty documents
• Network diagram and branch access setup
• WhatsApp contact for monthly care and reporting problems