Process

How Manar Builds

Every deployment follows the same five steps. Train, deploy, handover, adapt, compound. The steps do not change from one industry to the next. What changes is the data, the hardware, and the specific job the model is trained to do.

01

Train

A deployment starts from a registered model already built for the general classification a client needs. That starting model is then trained further on the client's own conditions. Its own cameras, its own documents, its own machines. Nothing generic stays generic past this step.

02

Deploy

The trained model installs on hardware the client owns, sized to what that specific model actually needs to do. A document search system and a sixteen camera vision system do not need the same hardware. Everything runs inside the client's building, offline if the client requires it, with no cloud dependency built into the arrangement.

03

Handover

A system nobody uses has failed, whatever the accuracy numbers say. The people who will actually work with it, the storekeeper, the shift supervisor, the records clerk, are trained on it before the engagement closes. Not a manual left on a desk. Handover is complete when your own staff can operate the system without Manar in the room.

04

Adapt

A model does not stay static once it is installed. It retrains as the client's own data grows, so it reflects the operation as it is today rather than as it was on the day it went live. This is scheduled with the client, not run silently in the background without their knowledge.

05

Compound

Accuracy on a client's own conditions improves with every day the model runs there. An organization that starts training on its own data now holds a record that a competitor starting later cannot shortcut or buy. This is the actual advantage of the approach, not a claim about the model being smarter in the abstract.

How a model is built

حِكمة

Hikma

Every model Manar deploys begins as an openly available foundation model. The strongest one available for a given domain, reviewed as the open field moves forward. What makes it useful to one specific client is what happens after that starting point.

It gets trained on the data of the exact domain and environment it will serve. Local conditions and local patterns no general training run has ever been shown. What leaves that process is not the same model that entered it. It belongs to the client, runs on hardware the client owns, and reflects a site no other model in the world has been trained on.

A typical engagement

A deployment usually starts with a site visit, not a proposal. Manar looks at the cameras, records, or systems already running before recommending anything, because the right model and the right hardware depend entirely on what is actually there. After that, training runs against real data from the site. Validation happens against the same conditions the model will run in once it is live. Installation and handover follow once the client has seen the results and is satisfied with them. Nothing about this timeline is fixed in advance, because no two sites start from the same place.

Model naming

Every model that gets trained is given a formal code before it is deployed: a family, a domain, and a method, in that order. A vision model trained for a security site is not the same registry entry as a vision model trained for quality control on a production line, even if both come from the same family. This is a discipline, not decoration. It means every model Manar has ever trained can be found, checked, and accounted for.

View registry Start a conversation