Support Desk Agent

Problem

Small support teams spend their day on repeat questions about late orders, damaged bags and subscription changes. Generic chatbots answer fast and invent policies.

Try it live

File a ticket as a customer, for example a torn bag. Switch to the agent console, read the cited draft, then approve, edit or escalate it. Deflection and accuracy counters update as you work.

Proves

Retrieval you can audit, human-in-the-loop design, and Django at production depth.

Signal chain

Stack

Django 5.2 · Django Ninja · Celery · Postgres + pgvector · Vue agent console · SSE

Engineering highlights

• A reply can go out automatically only when every claim in it cites a policy passage or a tool result.
• Hybrid keyword and vector search, with recall measured on a labelled set in CI.
• The policy corpus is embedded once at build time, so a ticket costs one injection check, one query embedding, one rerank and at most three model calls.