Muhammad Arshad

Muhammad Arshad

Founder of MAGSpace · Lahore

I've been building software since 2007. Now I build AI agents that can do real work inside a business, and that ask or stop when they should.

Why I started

AI can give an impressive answer. Can it do the work?

My first engineering job started in 2007. I spent most of the years since on telephony and high-availability systems. In that work, a call that guesses its route is not a small bug, it's an outage. That left me with little patience for software that is only approximately right.

Sources: LinkedIn profile · LinkedIn post, 16 Jan 2026

I joined BITLogix as a senior software engineer in 2009 and grew from there through architecture and R&D to EVP Engineering, which I left on 30 October 2026. Along the way I founded TechHatch in 2021 to mentor graduates through real projects. Since 2023 I've also led MisGenie as CEO.

Source: LinkedIn experience

When AI agents arrived, I saw the old problem again: confident output that is sometimes made up. In my own tests, a leading model invented a date in 3 of 39 runs. Checking the required fields in code before any tool ran brought that to zero. In 50 code edits I tracked, 16 needed no model at all.

Sources: X, 5 Nov 2026 · X, 6 Nov 2026

In November 2026 I left my job to work on that full time. I start from the business problem, not the model. The work has to fit how a company already runs, respect who is allowed to do what, and keep a person accountable.

Source: LinkedIn About

What I'm building

MAGSpace

MAGSpace is where I build products that let AI agents do business work while people stay in charge. An agent takes only an action it's allowed to take, asks when a detail is missing, and stops when nothing fits. Changes such as refunds wait for a person to confirm. Each product site says plainly whether it's live, in beta or in early access.

themag.space

What's live now

Products you can open today

Status labels as shown on themag.space and each product site.

Research · September 2026

Exact signals in whole numbers

Medical images and audio start as whole numbers: pixels and samples. I wanted to know whether they could be represented without floating-point rounding and still be recovered exactly. In September 2026 I published five preprints on Zenodo that prove when this kind of representation over binary-power rings can be uniquely recovered, and the verification code is public.

  1. Quantinion Simplex Cubic Decomposition over 2-Power Rings: Global Identifiability from Binary Contractions and Exact Bit LiftingDOI 10.5281/zenodo.22850767
  2. Exact Free Capacity of Symmetric Rank-One Cubic Dictionaries over Dyadic RingsDOI 10.5281/zenodo.22850810
  3. An Eight-State Operator and Transpose Geometry over Dyadic Rings: Trace-Zero Modules, Mixed-Radix Addressing, and Tensor-Lift KernelsDOI 10.5281/zenodo.22850830
  4. Dimension-Free Identifiability of (n+2) Symmetric Cubics over Dyadic RingsDOI 10.5281/zenodo.22850834
  5. Koszul–Sidon Rigidity for (n+3) Cubic Decompositions over Dyadic RingsDOI 10.5281/zenodo.22850842

Verification code Research hub ORCID 0009-0002-1314-0494

Writing

Recent posts

All writing →

How to reach me

Start with the problem

If you have a business problem that AI might help with, write to me and tell me about the problem first.