Industries
Manar maintains a registry of seventeen models across six families. Every deployment starts from one of them, trained further on the specific site that will run it.
Most institutions in Pakistan hold years of files that nobody has the hours to search properly. Board policy, past exam papers, student and staff records, government notifications, contracts, service rules. All of it sits in filing cabinets or scattered folders, in Urdu, in English, sometimes both in the same file. A Sijill model reads that archive and answers questions against it directly, citing the actual document a claim came from.
Cameras are already running in most of these buildings. Almost nobody is watching all of them, all the time. A Simah model watches instead, trained on the specific site it covers rather than a generic dataset of unrelated locations. A retail floor gets theft and empty shelf detection from cameras already installed. A plant or office gate flags an unrecognized face without needing a guard staring at a monitor all shift.
A textile floor in Faisalabad or a sportswear line in Sialkot produces more data in a shift than any manager can review. Cameras, machine sensors, and dispatch records all run separately, telling three different parts of the same story. A Simah model reads a weave or stitching line for defects as they appear. A Tanin model reads vibration and machine telemetry, so a quality issue and a machine fault get connected instead of investigated separately.
Stock and access in a lot of distribution hubs are still reconciled by hand, against records that are already a day behind by the time anyone checks them. A Simah model verifies stock against dispatch records without a manual count. A Sijill model reads access logs so who entered which bay and when is answered in seconds, not by pulling paper logs.
A single temperature excursion in a mango export shipment or a vaccine cold chain can spoil the entire consignment, and by the time it is noticed manually, it is usually too late to act. A Tanin model reads pressure and temperature sensors already installed and flags a drift the moment it happens, not at the next scheduled check.
A cotton holding or a mango orchard often has decades of yield sitting in a paper register that nobody has turned into anything usable. A Sijill model reads that register directly, so decisions about what to sow and where are based on the holding's own history rather than a provincial average that may not reflect one specific plot.
Patient records, prescriptions, and clinical notes in Pakistani clinics are often kept in a mixture of English and Urdu, typed and handwritten, structured and not. A system trained on a clinic's own archive can read that mixture back, answer questions about a patient's history, and surface the right document when a doctor needs it, without sending a single record outside the building it was written in.
A remote compression station or a pipeline segment runs sensors continuously. Pressure, temperature, flow rate, vibration. Almost none of it is watched in real time because the sites are remote, the data volume is high, and the staff is not there. By the time an anomaly is caught manually, the equipment has already been under stress for hours or the product has already been lost. A Tanin model reads that telemetry directly on-site, without sending it anywhere, flagging a drift the moment it moves outside the expected range. No internet connection required. The intelligence stays at the station.
If the data and the systems exist, a model gets trained on the operation specifically, whatever it is. A banking back office, a utility, a port, a transport company. The starting point is the same Hikma method used everywhere else on this page, scoped to whatever the work actually is.