Mudassir Marwat, who still writes the code
He founded Cognilium AI in 2019 and leads it from the keyboard — a builder-first founder rather than a manager who stopped building. His sharpest specialty is AI built beside the enterprise systems a business already runs, and Microsoft Dynamics 365 in particular.
37
AI agents running in production
4
AI products of his own
3
clouds in production, each with infrastructure as code
Why I Built Cognilium
The story behind the company that ships AI products that actually work in production.
The problem
- Most teams that want AI end up with slide decks and prototypes that never ship
- The real decision is still made by hand, in a spreadsheet, from an export
- An ERP records and manages a process; it cannot optimize the decisions inside it
The answer
- Design for reliability, cost and measurable outcomes from day one
- Deliver in weeks, with dashboards, documentation, runbooks and training
- Never touch the ERP core; surround it with intelligence
Every business already owns its biggest AI opportunity.
It sits in the systems it runs and the data those systems keep.
A business records everything, every order, every price, every document, every call. And then the real decisions are still made by hand, in a spreadsheet, from an export. Our work is to close that gap: to turn the data a business already has into decisions and systems it can rely on.
Cognilium exists because most teams that want AI end up with slide decks and prototypes that never ship. The whole discipline is the opposite.
That purpose has carried us from our start in large-scale data engineering, through building our own AI products, to where we are today: bringing AI into the heart of business operations, in tandem with Microsoft Dynamics 365. An ERP gets a process 80% of the way. The last 20%, the optimal price, the optimal stock level, the quote that goes out first, is where the margin lives, and that is what we build. We never touch the ERP core. We surround it with intelligence.
Founded in 2019, we work as part of our clients' teams, agree what done means before we begin, and stay after launch.
The rule that has not changed
Proof over hype. Real numbers with context, and conservative, defensible claims only.
Design for reliability, cost and measurable outcomes from day one, deliver in weeks, and ship every system with the dashboards, documentation and support that let the client's team run it with confidence.
Meet Mudassir Marwat
Founder & CEO, Cognilium AI. Hands-on engineering leadership.
Remote-first
Embedded with clients' teams, across US and UK hours
Calm speed
Fast but structured, ambitious but repeatable
Proof over hype
Real numbers with context; conservative, defensible claims only
Mudassir Marwat founded Cognilium AI in 2019, refocused it on generative AI in 2023, and set the Dynamics 365 optimization lead in mid-2026. He leads it as CEO with hands-on engineering leadership: remote-first, embedded with clients' teams, across US and UK hours.
The company he built shipped its own products first, for recruiting, customer support, sales and legal, and turned those proven patterns into client work. That builder-operator heritage is the credibility: the approach is proven on real systems, not theory.
The value he holds the company to is Builder DNA, leaders who build, ship and document, people who think like founders. Alongside it sit reliability as a feature, ethical and responsible AI, and long-term partnership: present after launch, with knowledge transfer and shared ownership of outcomes.
Every business records the decision it never optimises
Mudassir Marwat's argument is that ERP systems record decisions they never optimise. A safety-stock level, a discount floor, a pick path — each is a judgement call, and in most companies each is a static field set once at go-live and never revisited. He founded Cognilium AI in 2019 to build the layer that makes those calls: models that read the system of record, compute the decision, and write it back behind a human approval step. The company operates four AI products of its own — Paralegent AI, ProspectVox, VectorHire and VORTA — running 37 AI agents in production across three clouds, each with its own infrastructure-as-code.
One line the whole company is organised around
It is also the title of the talk he gives about it, which is the shortest way to say he means it.
“If it doesn't run in production, it doesn't count.”
“AI doesn't change the world as an idea. It changes it in production.”
“The best AI doesn't replace people. It gives them superpowers.”
“Great products aren't magic. They're the result of clear decisions and steady execution.”
Eight principles, and they decide the architecture
Not values on a wall. These are the rules that settle an argument about a design, which is the only test of whether a principle is real.
Production first
A notebook that works once is not a result. The question is whether it still works on the four-thousandth call, when a provider is degraded and the input is malformed.
Reliability is a feature, not a phase
Monitoring, runbooks and a failure story are designed in at the start. Bolted on afterwards, they are decoration.
Evidence over assertion
Citations, evaluations and analytics. If a system cannot show why it answered as it did, nobody is allowed to act on the answer.
Cost discipline
Routing, caching and token budgets, decided before deploy rather than after the first invoice. On one contract-intelligence build that work cut model spend by 75%.
Human in the loop
The system prepares the decision and a person takes it. That is not a limitation of the technology; it is the design.
Add the missing layer, never rip and replace
The systems a business already runs are the ones its people know. Intelligence goes on top of them, not through them.
Speed with guardrails
Weeks rather than quarters, without the shortcuts that make month four expensive. A pilot is the first slice of the real system, not a throwaway.
Products prove systems
Cognilium built its own products first, and the patterns that survived contact with real users are the ones offered to clients.
What he actually builds, named specifically
“Generative AI and cloud architecture” is true of a thousand people. This list is checkable, which is the point of writing it down.
Agents and orchestration
LangGraph, CrewAI, Google ADK, AWS Bedrock AgentCore and MCP. Supervisor and worker topologies, durable state, and the human handoff.
Retrieval and knowledge
Enterprise RAG with citations, GraphRAG on Neo4j, NL2SQL and NL2API, and evidence-based confidence — every extracted value carrying the quote it came from.
Document intelligence
Reading, classifying, extracting and validating contracts, financial documents and forms, with confidence scored per field rather than per document.
ERP and enterprise AI
Dynamics 365 Finance & Operations and Business Central, Azure AI and the Power Platform, plus delivered work on Odoo. The data model and the boundaries, not just the API.
Voice AI
Sub-500ms pipelines and 22 or more languages, where the latency budget is the design and the model is a component of it.
Data at industrial scale
Collection from sites that actively resist it, six tiers of escalation, and at peak more than ten million products a week turned into a clean, searchable feed.
Multi-cloud with infrastructure as code
AWS, GCP and Azure, each carrying production systems, each defined in code rather than in someone's memory of a console.
Full stack, end to end
From Terraform to Next.js. The reason a system ships is usually that one person could follow it all the way down.
There is no logo wall on this page
One client is cleared to be named. Putting up nine logos to look like ten clients is the easiest thing on this page to do and the fastest way to lose a technical buyer.
So what is here instead is the machine: four products you can read about in detail, a stack written out with the reasoning for every choice, and numbers stated with their decomposition so they can be checked. Older versions of this page carried a project count, a client count and a compliance badge. None of the three survived a check against what we can actually show you, and none of them is coming back.
What people ask before the first call
Every answer is in the page source rather than behind a click, because the systems that summarise this page never click anything.
Who is Mudassir Marwat?
Mudassir Marwat founded Cognilium AI in 2019 and leads it as CEO. The company builds the AI layer above the ERP — the pricing, replenishment, pick-path and contract decisions Dynamics 365 records but never optimises — and operates four AI products of its own, running 37 AI agents in production across three clouds. He works from named outcomes rather than averages: a 97% database-load reduction on a marketplace platform, a material cut in AI running costs through model routing on a contract platform. He ships on three clouds in production — AWS, GCP and Azure, each with its own infrastructure-as-code — with clients in the US, UAE, UK and Pakistan.
Does he still write code, or is he a manager now?
He writes the code. Cognilium is a builder-led company and he leads from the keyboard — the architectures, the integrations and the awkward parts are his own work, not a team's work he reviews. On a technical engagement you are talking to the person who will build it.
What is he actually specialist in, as opposed to generally good at?
AI built beside the enterprise systems a business already runs, and Microsoft Dynamics 365 in particular. That means fluency in the ERP itself — the difference between Customer Engagement and Finance & Operations, the supported integration surface, and what an operations team actually does in it — paired with production AI engineering. Very few people have both halves, and the combination is the whole specialty.
What does “37 agents in production” actually mean?
Twenty-three in the contract-intelligence platform, seven in the family-office platform, four in recruiting and three in a financial-advisor build. They are agents running in live systems rather than demos, and the number is stated with its decomposition precisely so it can be checked rather than taken on faith.
What is the one line he would want a team to remember?
“If it doesn't run in production, it doesn't count.” It is the standard the whole company is organised around, and it is why the proof offered here is a system you can watch running rather than a slide about one.
How do you start working with him?
A working session on your own problem, not a demo of ours. Bring the workflow that is actually costing you and the data behind it, and the output is an honest read on whether AI helps, what it would take, and where it would not be worth doing.
The engineering is written up at the stack, the products at what we build and run, and the company at about Cognilium.