Exception Desk

• 

  Recorded fictional live result: quoted model evidence and fixed safety checks route an urgent case to human review.

Why we built it

A small service business can receive an urgent safety concern, a routine request, and a message missing key details in the same inbox. Exception Desk explores how AI can turn each note into a checkable handoff without letting the model make the final operational decision. It is an ImpactLayer work sample, using fictional cases throughout.

What it does

An operator enters a fictional customer note. A Cloudflare Worker sends it to the NVIDIA Nemotron-3-Nano-30B-A3B model through Nebius Token Factory. The model extracts facts, quotes supporting source text, identifies missing details and suggests a reply. Separate fixed rules check hazards, service area and explicit contact consent. The screen shows the model's suggestion as untrusted, the rule checks, the final human-review route and an unsent draft. It never sends a customer message or creates a booking.

How we built and tested it

The browser UI, Node.js policy module and Cloudflare Worker share the same routing logic. A D1 counter caps live model attempts at 20 for the lifetime of this demo; a per-IP rate limit protects the public endpoint. The Nebius API key lives only as a Worker secret. Twelve automated tests cover urgent text overriding a routine model classification, invented consent and location not clearing checks, safe fallback, and the live-request adapter.

On 29 September we verified a live model-backed run: a fictional gas concern returned HTTP 200, quoted the source evidence and routed to ESCALATE, with no message sent and no booking created. A dripping-tap example routed to clarification. The recorded safety run linked below preserves that real result without making another billable model call.

Challenges and learning

One model response exceeded its generation budget; another contained no usable JSON. We constrained the requested output and made parsing failure route safely to a person while source-based hazard checks still run. The lesson is that useful AI implementation needs visible evidence, bounded usage and rules that remain trustworthy when model output is incomplete. This is a demonstration, not a deployed customer system or a claim of measured business results.

Try the fictional demo: https://impactlayer-exception-desk.impactlayer.workers.dev/

Recorded live safety run: https://impactlayer-exception-desk.impactlayer.workers.dev/?replay=safety

ImpactLayer: https://impactlayer.co.uk/

Built With

• cloudflare
• d1
• javascript
• nebius
• nemotron
• nvidia