Skip to main content
Concept build · AI & automation

Relay — an AI support-inbox automation console

A working AI console that triages a support inbox — classifying messages, extracting structured fields, recommending an action, and drafting a reply. Explore the live demo.

Service
ai-automation
Concept build
Live interactive demo
Reading time
3 min

This is a concept build, not a client engagement. We built Relay ourselves to show how TechSpireX approaches an AI-automation product: the part that is actually hard is not calling a model — it is the interface and the guardrails around it. Every triage decision in this demo is pre-authored, not live inference, so the build is self-contained and deterministic. Open the live console and work through the inbox yourself.

→ Open the live Relay demo

The idea

Support inboxes are where a lot of a team's time quietly disappears. Most incoming messages are variations on a handful of intents — a billing question, a bug report, an onboarding how-to — and a good automation layer can handle the routine ones end to end while making sure the genuinely sensitive ones reach a human fast.

Relay is that layer, as a product surface: an inbox where every message arrives already triaged.

What it does

For each incoming message, Relay shows:

  • Classification — category, sentiment, intent, and urgency, with a confidence score.
  • Structured extraction — the fields that matter pulled out as tidy chips (invoice IDs, amounts, VAT numbers, affected endpoints, renewal dates).
  • A recommended action — auto-resolve, route to a team, or escalate to a human — with the reasoning for it.
  • A drafted reply you can edit, regenerate for an alternate tone, and approve.

The console streams the drafted reply in as it "generates," runs a brief analysis state on each message, and lets you filter the queue, regenerate drafts, and approve or escalate — updating the message's status live.

The decision that actually matters: when not to automate

The interesting design work in an AI automation tool is the boundary. A duplicate charge under the refund threshold is safe to auto-resolve. A price-driven cancellation before a renewal date is not — it gets escalated to a human with a suggested retention offer, never auto-cancelled. A sudden spike of API 429s in production is paged to on-call, not answered with a template.

Relay encodes that judgment explicitly: the recommended action and its rationale are first-class in the UI, so the human stays in control of anything with real consequences. That is the difference between an automation a team trusts and one they turn off in a week.

Engineering notes

  • Accessible, controllable UI — real form controls, keyboard-operable, a streaming reveal that collapses to instant text under prefers-reduced-motion.
  • Deterministic by design — no external calls, no keys, safe under a strict Content Security Policy; the same data renders identically on every build.
  • A distinct product theme — a dark automation console, separate from this marketing site, sharing a design language with our other concept build, Meridian.

The stack

Next.js (App Router), React with strict TypeScript, and Tailwind CSS — a single client island rendered inside an otherwise static site. Wiring this to a real model (hosted or via an API) is a small, well-bounded change; the product design is the part shown here.

The honest caveat

Relay does not call a real model and has no real inbox — it is a demonstration of how we design and build an AI-automation product, not a claim of a delivered client result. Real, attributable case studies will be clearly marked as such when we publish them.

→ Open the live Relay demo