The sector decides what has to be true
The models are the same everywhere. What changes is the systems we have to reach, the words the retrieval has to know, and what your reviewer will insist on before anything ships. Cognilium AI is an AI company, and below is where its systems have actually been built — named honestly, and separated from what we can build but have not yet delivered.
8
verticals with a delivered system behind them
3
with a full engineering page of their own
1
named client we are free to name: Dyco Parts
Eight verticals, one delivered system behind each
No logo wall and no invented practices. Role and vertical descriptors only — Dyco Parts is the one client we are free to name. Several of these are our own products, which is why we can open them up rather than describe them.
Financial services
An AI chief of staff for a family office: seven agents, a twelve-step document-intelligence pipeline across 25 document types, and a temporal knowledge graph where every extracted value carries its source quote.
See the work →Manufacturing and automotive
A delivered, reviewed engagement bringing generative AI and optimization work inside an automotive-parts manufacturer's Dynamics 365 Supply Chain environment.
See the work →Distribution and ERP
Dyco Parts, a truck-parts manufacturer: a C-level decision assistant over an Odoo ERP, documents, e-commerce data and scanned invoices, delivered in Slack.
See the work →Legal
Paralegent AI, our own product: 23 agents reviewing a contract against the firm's own playbook, inside Word, in 5 to 10 minutes.
See the work →Marketplace and e-commerce
A monitoring platform re-architected onto a distributed cloud: around 2,000 users across 16 marketplaces, 3,000 checks every two minutes, with a 97% database-load reduction from caching.
See the work →Education
A writing co-pilot embedded in the learning platform a K-12 education company already ships, grounded in their own curriculum, with a judge scoring every lesson before it reaches a classroom.
See the work →Automotive SaaS
Senior architecture across three interconnected platforms — auction, appraisal and dealer websites — on a shared multi-tenant core, with API governance and technical leadership.
See the work →Recruiting and HR
VectorHire, our own product: resume, professional-network, code-host and live voice-interview agents screening candidates in parallel.
See the work →Three things, and none of them is the model
“Do you know my industry?” is really three questions. These are the three, and they are what a first call should be about.
The systems we have to reach
An ERP in distribution, a historian and MES on a plant floor, a core banking platform in finance, a document system in legal. The integration surface is the part that decides whether a project is six weeks or six months.
The vocabulary the model has to get right
A distributor's customer says 'those valves'; a plant says BPFO and Hotelling T-squared; a bank says pacs.008. Retrieval that does not speak the sector's own words returns plausible answers to the wrong question.
The posture the reviewer will demand
Regulated sectors decide the topology before the model: where inference runs, what leaves the network, what the audit log has to prove. That conversation happens first, or it happens at the end and kills the project.
Three sectors written for the engineer
Named protocols, named standards, named integration points. Two of these carry delivered work behind them; the third is capability, and is labelled that way on its own page.
Financial services
Sub-15ms fraud inference targets, reason codes on every automated credit decision, AML entity resolution, and topologies that pass a model-risk review.
Read the engineering →Manufacturing
OPC-UA and Sparkplug B pipelines, ISA-95 modelling, golden-batch deviation under GAMP 5, and inference that never writes to the PLC.
Read the engineering →Marketing
Bayesian marketing-mix modelling calibrated by geo-lift, server-side conversion APIs, and clean-room identity that is consent-aware at the router. Capability, not yet a delivered engagement — the page says so.
Read the engineering →
Common questions
Asked before the first project, answered honestly
Financial services, manufacturing and automotive, distribution and ERP, legal, marketplace and e-commerce, education, automotive SaaS, and recruiting and HR. Each of those has a system actually delivered behind it, and several are our own products — which is why we can show you the inside of them. We also publish a marketing and advertising page: that one is capability rather than delivered record, and it says so.
No, and it is a fair question to ask directly. The mechanisms transfer: retrieval over your own documents, agents that use your tools, document intelligence with evidence, pipelines that survive real load. What does not transfer for free is the vocabulary and the integration surface, so a first project in a new sector is scoped with that in mind rather than priced as though we already knew it.
By deciding topology before models. Where inference runs, what leaves your network, which credentials are scoped where, and what the audit log has to be able to prove — all agreed with the people who own your security review before anything is built. Our architectures are aligned with SOC 2, ISO 27001, GDPR, HIPAA and PCI-DSS practices; we do not hold, and do not claim, formal certification in any of them.
That is usually the majority of the work. We read from ERPs, MES and historians, core banking platforms, document management systems, applicant tracking systems and CRMs — and we are deliberate about what we write back. On the plant floor we never write to the PLC; beside an ERP we write only what we created.
The typical time to a first working system is about three weeks, and we say up front when something will not work. Beyond that we quote against your actual integration surface rather than a blanket estimate, because the difference between reading a documented API and reading a historian nobody has audited is months, not days.
Least-privilege access, encryption in transit and at rest, role-based controls and audit logging on every request, with deployment inside your own cloud account where the sector requires it. Where a regulated environment needs it, the whole model lifecycle runs inside your network with no public egress.
Tell us what your sector runs on
This is our AI engineering practice
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.