Customers
Not a wall of marks. Every customer here comes with what was built and where it runs — from a trading house's analyst portal to a bank's loan assistant. Two are written out in full, including the one still in delivery.
The stories
What each one built and where it runs. The first two open into a full story.
Trading house · M&A
Home improvement · HR & call center
Retail & commercial banking
M&A advisory
Universal banking · State bank
Pension fund administration
Fintech · Digital payments
Medical education · Pharma & biotech
Higher education · Student services
Telecom · Investor relations
Software · HR operations
Consumer tech · Internal HR
Story 1 of 12: Itochu
What the full stories show
The interesting part of a customer story is rarely the industry. It is the shape of the constraint — and in the two written out in full, it is the same three things.
Both deployments sit inside the customer's existing cloud tenant, under their own identity provider and their own security posture. Neither organisation was willing to route the material through a vendor's infrastructure, and neither had to — which is why the deployment question was settled early rather than becoming the thing that killed the project.
Filings, agreements, internal policy, reference documents. In both cases the value is not that a model can write; it is that a specific question can be answered from a specific corpus with the passage cited. A general assistant would have been faster to deploy and would have answered a different question.
Both systems choose a model per unit of work rather than being built against one provider. Over the life of either engagement, model availability, capability and pricing all moved — which is the argument for the property, and it is easier to make retrospectively than in a procurement meeting.
The first one is asked sceptically, and fairly.
Because these are the names cleared for a public page. A customer agreeing to be a reference on a call is not the same as agreeing to be printed, and the previous site carried names nobody could show had been cleared. There are more deployments than there are slides here; the constraint is clearance, not count.
Ask. Reference conversations are arranged one at a time, with the customer's agreement each time, and they are far more useful than a case study because you get to ask the awkward questions. If there is nobody appropriate to introduce you to, you will be told that rather than pointed at a page.
Not here, because the ones that existed could not be sourced. The previous version of this site carried outcome numbers attributed to named organisations — including one under an active takedown — and none of them traced to a measurement anybody could produce. What is on these pages instead is a description of what was built and an honest statement of what stage it is at. If a number appears here later it will be because someone measured it and the customer approved printing it.
Talk to a NeuralSeek expert about how it fits your stack, where your data has to live, and the governance your auditors already expect.
