We design, build and run the systems your business depends on, then stay to keep them running. Four AI products in production, three clouds, and a typical three weeks from start to something live.
Trusted by enterprises
Built on leading technologies
37
AI agents in production
2019
founded
3
clouds, each with IaC
Every order, every price, every stock level is already in your system. Then someone exports it to a spreadsheet to work out the answer, because the ERP stores the rule and not the judgement. That export is the gap. It is where your margin gets decided, by hand, every week.
rebuilds safety stock in a spreadsheet every month, because the field in the system was set once at go-live and nobody has revisited it since.
approves discounts by rule of thumb, because the system enforces the price that was typed into it and has no opinion about the one that was right.
re-types a customer's emailed mess into an order, line by line, and the fastest quote out the door usually wins the business.
“Pulling any information into excel is extremely time consuming and takes a lot of manipulation.”
None of that is a failure of the ERP. An ERP is built to record a transaction and enforce a rule, and it does both well. Computing the better answer is a different job, and it is the one we do.
We work in two lanes and they want different things from this page. One is already running Microsoft Dynamics 365 and wants better decisions out of it. The other is building an AI product and needs it engineered properly. Pick the one that sounds like your week.
For operations and IT leaders at manufacturers, distributors and retailers whose ERP is live and working. We build the optimization apps that compute the pricing, stocking, picking and contract decisions it records but never calculates. They run beside Dynamics, on the stack your IT already governs.
AI in the ERP Building a productFor founders, CTOs and heads of product. Agents that do real work, answers drawn from your own documents and data, document intelligence and voice. This is the practice that built our four products, and it is where most of our engineering hours still go.
The capability mapIt reads what Dynamics already holds, computes the decision outside the ERP core, and writes back only the answer, behind a human approval step. Four of them, each named after the decision it makes. One is shipped and in daily use; the other three are built against your own data on request.
Dynamics applies the price you typed into it. This computes the one your history says you should have typed, per customer and per line.
Placement derived from your own order history, so pickers walk less. Travel is roughly half of order-picking time, and picking is the largest single component of warehouse labour cost.
Reviews a contract clause by clause against your own playbook, inside Microsoft Word. Shipped, in daily use, and sold as Paralegent AI.
Safety stock and reorder points computed from your demand history and written back where planning already reads them.
The boundary, before you ask
The ERP core is never modified. We read through Dynamics' own supported surface, compute outside it, and write back exactly what we created: the recommendation, and the record of who approved it. Nothing else. Your implementation partner's work stays exactly as they built it, and they keep it.
Our focus right now
In most warehouses, roughly half of picking time is spent walking rather than picking, and the walking is driven by where inventory was placed, often years ago and never revisited. We read your order history and tell you what to move, inside the Dynamics 365 you already run, so more lines go out the door with the same crew.
Two teams quote the same Dynamics project and one of them has read the documentation. These are the surfaces we build against and the detail that decides whether a first sprint produces anything. If your team knows this already, we will get along.
Customer Engagement runs on Dataverse and the Power Platform. Finance and Operations is the ERP core, with a different data layer, different APIs and a different environment model. Asking for the wrong environment costs weeks before anyone touches data.
OData data entities for master data, inventory and locations. Custom services for logic no entity covers. Business Events to react to receipts and put-away as they happen instead of polling. All documented, all still there after an update.
A Microsoft Entra app registration has to be registered inside Finance and Operations and mapped to a service role before one API call will succeed. It is one screen, it is in the documentation, and not knowing it is the most common reason a first Dynamics sprint produces nothing.
Business Central online publishes a wide set of standard endpoints, and a few things a real workflow needs are not among them. Item references and price lists come up first. The answer is a small companion extension covering exactly that gap, installed once, rather than a sync that drifts.
Power Platform, Dataverse and Azure, in your tenant, with the identity model and data residency your security team already signed off. Nothing we add asks a security review to start over.
The same discipline against a different published contract: read through the platform's own supported surface, compute outside the core, write back only what we created. The optimization patterns are not Microsoft-specific. The integration work is.
We have no Dynamics customer references yet and we say so. What we have instead is four AI products we built and run, and client systems whose mechanisms we can describe in full. Each line below pairs a moment you will recognise with the thing we actually built to close it.
You re-type a customer's emailed mess into a quote, line by line.
23 agents read a contract inside Microsoft Word in 5 to 10 minutes: 12 scoring the categories, 11 analysing the clauses that matter. A legal team used to spend hours on the same document. Routing cut the LLM calls by 75%, so the same answers cost a quarter of the spend.
Paralegent AI, our own product
You do not trust the number the system hands you, so you check it.
Every extracted value carries the exact quote it came from and the page it sat on. On our own document platform, tested across 25 document types: 100% classification, around 97% quote accuracy, and scanned documents lifted from around 12% usable to around 94%.
FamilyOffice.ai, our platform. Figures are project test results
You have been promised eight weeks and given twenty-two.
We write roughly 33,000 lines of checks against roughly 17,600 lines of application code, and 2,092 of them run in under ten seconds. The quality judge fails open by design, because a judge that goes down must never take the product down with it.
How we run every engagement
The last system slowed to a crawl once real volume hit it.
A caching layer cut one platform's database load by 97%: a TTL state machine, five atomic Lua scripts, and 250 lines of cleanup code deleted outright. It carries 3,000 tasks every two minutes, around the clock, across 16 marketplaces, scaling from 2 workers to 40 on queue depth.
A delivered engagement
Your team already exports to a spreadsheet because search never finds it.
For one manufacturer we put a multi-agent system over their ERP inside Slack, where the executives already were: agents for structured queries, for documents, and for reading scanned invoices, tied together over a knowledge graph.
Dyco Parts
37
AI agents running across four of our own products
3
clouds in production, each defined in infrastructure as code
2019
founded, with clients in the US, UAE, UK and Pakistan
What we cannot show you yet
No Dynamics customer reference, no AppSource listing and no Microsoft partner designation. That work is new. What we can do is show you a system running, walk you through the architecture, and build the first one against your own data so you are judging our work and not our slides.
One workflow, four to six weeks, one success metric agreed before we start and read out weekly. No throwaway code: the pilot is the first slice of the real system. A working proof of concept lands in days and something in production in about three weeks.
We scope before building, and we check your data can answer the question before we write the proposal. If it cannot, we say so then.
Architecture and acceptance criteria agreed up front. What done means is settled before anyone starts building toward it.
Milestones tied to working software, demos every week, replies within hours. Working proof of concept in days, something in production in about three weeks.
Every deployment ships with documentation, dashboards, runbooks and training, and we stay after launch.
Engagement models
Fixed-price, hourly, or retainer.
It is real work and it is where our four products came from. But what Cognilium leads with is narrower: optimization apps that run in tandem with Microsoft Dynamics 365, computing the decisions the ERP records but does not derive — the optimal price, the optimal pick path, the optimal stock level. See the optimization apps · How we build inside the ERP.
What we do
From AI development to full-stack solutions - we build enterprise-grade systems that scale
AI built inside Microsoft Dynamics 365. Pick optimization, vision to ERP, knowledge assistants.
Azure-native
Conversational bots, support agents, domain-tuned models with your data (RAG, LangChain)
Production-ready
We turn ideas into full-stack SaaS apps — React + Django, MERN, shipped and deployed
Full-stack
Vector DBs, embeddings, OpenSearch, Glue, BigQuery — we build the pipeline end to end
Retries and back-pressure
Kubernetes, Docker, GitHub Actions, CI/CD pipelines for AI workloads
Auto-scaling
50+ cloud VMs scraping millions of product and price points daily
50M+ Data Points
Strategic AI implementation planning, technology roadmaps, and executive advisory services
C-Suite Focused
Success stories
Systems that reached production and stayed there — enterprise search, agent orchestration, document intelligence.
Five steps from a first call to something running in your stack, and the two people you will be dealing with. No account manager in between, and no handover to a team you have not met.
Step 1 · Outlook booking
A 30-minute call with an engineer who has shipped this work, not a sales rep.
Step 2 · Straight answers
Bring the stack, the constraints and the parts you are unsure about. You get straight answers, including where we think the idea is wrong.
Step 3 · Dynamics 365
We agree what is in, what is out, and whether this needs one engineer or a full squad.
Step 4 · LangChain, OpenAI, Claude
An architecture you own: the models, the data flow and the integration points into the systems you already run.
Step 5 · AWS, Azure, Kubernetes
Engineers embed with your team and ship to production in your stack and your sprint cadence.
One engineer or a full squad, scoped to the work rather than sold as a package. The first call is thirty minutes with somebody who has shipped this, and you will get a straight answer including where we think the idea is wrong.
Mudassir Marwat founded Cognilium AI in 2019 and still leads from the keyboard. A decade building software, beginning in data systems engineering. His sharpest specialty is AI built beside the enterprise systems a business already runs, and Microsoft Dynamics 365 in particular.
“If it doesn't run in production, it doesn't count.”
He pairs AI engineering with working fluency in the ERP itself: its data model, its boundaries, and what an operations team actually does inside it. On the first call you are talking to him or to an engineer who has shipped this work, not to a sales team.
Every answer is in the page source rather than behind a click, because the systems that summarise this page never click anything.
We build AI optimization apps that work in tandem with Microsoft Dynamics 365. Dynamics is your system of record; we are your system of intelligence. It manages pricing, warehouses and inventory; we optimize them with data science. We also run a separate AI engineering practice building production agent, retrieval and data systems.
We do AI, and we do it inside the ERP you already run. Your partner put Dynamics in and keeps it running, and we make what they built worth more: the complementary apps that make the optimal call on top of it, the last mile your ERP records but cannot optimize. We don’t touch your ERP core; we surround it with intelligence.
Dynamics enforces the rule and stores the data, which gets a process about 80% of the way. The last 20% is the decision itself: the optimal price, the optimal pick path, the optimal safety stock. Those need data science, not configuration. Our app computes the recommendation and the process owner acts on it inside the system they already use.
The lead plays are pricing and discounting, warehouse pick-path and slotting, contract and procurement review, and demand and inventory planning. We pitch across roughly twenty ERP domains in total. Working demos exist and we build each app on request.
We run in tandem with Dynamics on your own governed stack — Power Platform, Dataverse and Azure — and ship via AppSource where appropriate. We read from and write to documented integration surfaces rather than modifying the ERP core, so your implementation partner's work and your upgrade path are untouched.
Cognilium AI is remote-first and works as part of our clients’ teams across US and UK hours. Clients are in the United States, the United Arab Emirates, the United Kingdom and Pakistan. Delivery runs over video calls, Slack and whatever project tooling your team already uses, and Dynamics work runs inside your own tenant rather than ours.
We built 4 production AI products: Paralegent AI for contract review with 11 AI specialists, ProspectVox for voice AI sales with 4 agents, VectorHire for AI recruiting with parallel screening agents, and VORTA for 24/7 AI customer support. These products prove our engineering depth.
Yes. We provide pre-vetted senior GenAI engineers who embed with your team within 48-72 hours. They work in your sprint cadence, your stack, your tools. Expertise includes LangChain, OpenAI, RAG, multi-agent systems, and voice AI. No long-term commitments required.
For Dynamics work: Power Platform, Power Apps, Dataverse and Azure, shipped via AppSource where appropriate, so everything stays inside your governed environment. For the AI engineering practice: LLMs (Claude, GPT, Llama, Mistral), vector databases (Pinecone, Weaviate, Qdrant, pgvector), orchestration frameworks (LangGraph, CrewAI, LlamaIndex) and AWS, Azure or GCP. All with monitoring, testing and deployment pipelines.
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.
