Demonstration study · AI
Signal Desk
A demonstration operations desk that drafts routine replies and routes requests, with a person approving anything that leaves the building.
TypeScript · NestJS · LLM APIs
Concept visual for this demonstration. Not a screenshot of client work.
Overview
Signal Desk is a demonstration study, not a client engagement. It shows how ZiyaCore would place a language model inside an operations workflow: summarize the incoming request, draft a response from approved material, and wait for a person before anything is sent.
Business challenge
The team answers the same classes of request all day. The knowledge exists, but it is spread across old threads. New staff take longer, and the answers drift.
Technical challenge
A model can draft quickly and still be wrong. Customer data should not be sent to a vendor by default. The product needs a record of what was suggested, what was edited, and what was actually sent.
Solution
A web desk backed by a NestJS service. Incoming requests are classified against a short set of types. A model drafts from approved source material on the server. A person edits and sends. The system stores the draft, the edit, and the outcome. Anything outside the known types is queued for a person with no draft presented as final.
Design strategy
The screen is a review desk, not a chat. The request, the sources, and the draft are visible at once. Sending is a deliberate action. Low-confidence and out-of-scope items are visually distinct from drafts that are ready for review.
Engineering implementation
Model calls happen on the server. The client never sees vendor credentials. Each run is logged with the sources used. The send action is a separate operation from the draft, so a generation cannot be mistaken for delivery.
Verified outcomes
ZiyaCore publishes performance results only when a client has agreed to share them. This page is a demonstration study, so there are no metrics here. Nothing on this page should be read as a client result, a testimonial, or a guarantee.
