Deskloop — AI-Assisted Support Triage
Link to open source: https://github.com/AnushkaSharma92/swytchcode-1
Description: Deskloop turns an incoming support email into a fully-handled case with a human still in control at every step. It pulls the email (from Gmail or a fixture inbox), classifies it with an LLM (category, confidence, key entities), searches an internal knowledge base for grounded context, and drafts a reply that only cites real KB articles — never fabricated ones. A support agent reviews the draft, optionally escalates to a Jira ticket (with related GitHub issues surfaced automatically), and approves sending via Resend or Gmail. Every step is logged to an audit timeline for full traceability. An integrated AI Assistant panel lets the agent ask follow-up questions about the case — classification reasoning, next steps, refining the draft — grounded strictly in that case's data, with guardrails against prompt injection from email/KB content.
Key features:
- Automated email classification (bug / billing / how-to / refund / unclear) with confidence + reasoning
- Grounded reply drafting — cites only real KB matches, refuses to guess
- One-click ticket escalation to Jira with related GitHub issue lookup
- Human approval gate before any email is sent
- Full audit timeline per case
- Conversational AI assistant for agents, scoped to the active case.
