Difae Research builds machine perception systems for defence and security.
Everything we build is designed to run at the edge: on constrained hardware, without a connection to anything.
Work
02
What we work on
We are an independent research company. We work on a small number of problems in detection, tracking, and autonomy, we build against real hardware budgets, and we publish what we can.
Two areas. The scope is narrow on purpose: these are problems where the failure modes are expensive and the interesting engineering is in the edges, not the average case.
- 01
Machine perception for surveillance
Detection, tracking, and classification from imagery, video, and other sensor streams. Most of the work is in the parts nobody demos: small objects, poor light, cluttered scenes, sensors that disagree with each other, and teaching a model to decline to answer instead of guessing.
- 02
Autonomy in degraded conditions
Systems that have to keep working when the network is gone, the position fix is wrong, and the operator is doing three other things. We are interested in graceful degradation: what a system does once its inputs stop being trustworthy, and how it tells a person it is no longer confident.
Constraint
Edge-first, by design
Both areas share one constraint that shapes everything else: the model runs on the device.
No inference in the cloud. No telemetry leaving the platform. No dependency on a link that someone else can cut. This is an engineering position and a security position at the same time. A system that never sends data anywhere cannot leak it, and a system that needs no network cannot be denied one.
Working this way means designing inside real budgets rather than treating them as an optimisation problem for later. Quantisation, pruning, and distillation are defaults, not a final pass. Power and thermal envelopes are inputs to model selection, not consequences of it. And a model has to be updatable in the field by someone who is not an engineer, because the alternative is a system that is correct on the day it ships and slowly stops being correct after that.
Research
Open research
Alongside commissioned work we run our own projects. Some of that work is not publishable and we would rather say so than imply otherwise. What we can share goes here: technical notes, negative results, and occasionally code.
The first notes will be published as the work reaches a point where there is something honest to say about it.
Responsibility
How we think about this work
Surveillance and defence technology is not neutral. We would rather state our position than leave it to be inferred.
A person decides.
We build systems that detect, classify, and recommend. We do not build systems that select or engage a target without a human in the decision. This is a line, not a preference.
We check who the work is for.
We assess the export-control classification of what we build before an engagement rather than after, and we decline work where we are not willing to stand behind the end use or the end user.
We describe what a system cannot do.
Performance figures are quoted with the conditions they were measured under. A model that works in clear daylight is described as a model that works in clear daylight.
Firm
Difae Research
Difae Research was founded in 2026 and is registered in Pakistan. It was started by Muhammad Awais, who builds and operates AiBayya, a production system merchants use to sell over WhatsApp.
We are small and intend to stay that way for a while. The work is done by the people you talk to.
We also build and operate commercial software.
AiBayya is an AI assistant for merchants selling over WhatsApp, built and run by the same team. It has nothing to do with the work above, but it is a live production system with real users, which is a different kind of evidence than a research note.
Contact
Get in touch
If you have a problem in detection, tracking, or on-device inference, describe it and we will tell you plainly whether it is something we can do. We say no often.
hello@difaeresearch.comPlease do not send classified, controlled, or otherwise sensitive material by email.