Practices
PMC DXB Practices Applied AI systems
Applied AI systems — PMC DXB, Dubai
I am Sumesh Paul Meledath. I build the systems that do a Gulf company’s administrative work — supplier invoices read out of WhatsApp, quotes drafted with VAT on them, WPS files checked before they reach the bank, customers answered in Arabic and English — inside the tools your team already uses. One principal, fixed scope, fixed price.
Two of those systems are running today. One of them answers enquiries on the front page of this site — you can put a question to it yourself before you ever contact me. I would rather you tested the work than read an adjective about it.
No pitch deck. Twenty minutes on WhatsApp or a call. You leave with a written list of what is automatable in your business — whether or not you hire me.
The work
Not “digital transformation”. These are the specific, repetitive tasks that consume a UAE SME’s week, and the shape of work this practice is built to take on. Your list will differ — establishing that is what the audit is for.
Every action is logged and human-reviewable. An agent you cannot audit is a liability, not an asset — that is the design constraint the whole practice is built around.
Proof
Client names are withheld under NDA; references on request. What follows is what I can show you without asking anyone’s permission.
Peter runs the front line of this website.
Mudeer is deployed at a Dubai trading company.
If you want the longer version of how this is done and where it fails, I write it up: AI automation for UAE businesses and the fixed-scope automation engagement behind Mudeer.
Deployment
Some of this work cannot sit behind someone else’s API, and in the Gulf that is a regulator’s question more often than a preference.
Where data cannot leave your jurisdiction, your building or your regulator’s reach, the model runs on owned hardware rather than behind someone else’s API. I develop and run this on my own Apple Silicon in Dubai — a 60-core M3 Ultra with 96 GB of unified memory, large enough to hold a frontier-class open-weight model resident and serve it without a single inference call leaving the machine. Where your own policy requires it, the same architecture is deployed on hardware you own and control.
That is the honest shape of sovereign AI at this scale: not a datacentre, but a private stack you own outright. If your requirement genuinely needs a datacentre, I will tell you on the first call.
Process
Four steps, and you can stop after the first one with something worth having.
Boundaries
I do not train foundation models and I do not build hardware — I buy it, run it, and know precisely what it will and will not hold. If your programme needs pre-training or silicon, I will say so on the first call rather than the fifth.
I am one principal, not an agency. That is the reason the scope is fixed and the price is quoted before the work: it is the only honest way to sell a single person’s attention. Where the work is genuinely larger than me, the answer is a referral, not a hire.
Questions
Fixed scope, fixed price, quoted in writing after the free 20-minute audit. I do not publish a rate card because the honest number depends on how many of your processes are actually automatable, which is what the audit establishes. You will have the price before you commit to anything.
The audit is 20 minutes. After that it depends entirely on scope, and the quote states the timeline alongside the price. A single process is not a six-month programme, and I will not sell it as one.
Only if you want it to. Where residency matters, the same architecture is deployed on hardware you own and control, so no inference call leaves your premises — see private deployment above. Where it does not matter, hosted models are cheaper and I will say so.
Yes, and Arabic-first rather than translated. Mudeer covers four languages with Arabic as the primary, and customer replies in Arabic are part of the standard task set. You can open the demo and check.
The ones you already run — accounting, CRM, WhatsApp, email and document storage. The design rule is that your team should not have to learn a new interface to get the benefit. If something in your stack genuinely cannot be integrated, that comes out in the audit rather than after the invoice.
Every action is logged and human-reviewable, and exception cases are routed to a person by design. That is a deliberate constraint: an agent you cannot audit is a liability, not an asset.
Next
You leave with a written list of what is automatable in your business, whether or not you hire me. That is the whole offer.
This practice sits under 01 Applied AI systems. The security, capital and mobility practices are on the main page, and who I am is a separate read.