Production AI support-agent SDK for an enterprise client
Built and operate a TypeScript AI agent SDK that connects an IT/ITSM provider's chat channel to a ticketing system with LLM-based classification, replacing a discontinued third-party low-code platform and handling real support traffic in production.
- Node.js
- TypeScript
- LLM-based classification
- Ticketing (ITSM) integration
- systemd
Context
An IT/ITSM provider serving multiple companies under an outsourced-support model needed its chat channel connected to its internal ticketing system, with an AI layer able to classify and route tickets — replacing a third-party low-code platform that had already been migrated once and was discontinued again within a year.
Challenge
Ticket categorization that didn’t scale (exact keyword matching only covered a handful of hundreds of possible categories), natural-language ambiguity in ticket-number references, and building something that runs reliably as production infrastructure — not a demo — including session resilience across restarts and multi-tenant credential routing, since the same SDK serves more than one internal bot.
What I did
- Rebuilt the business logic as an in-house Node.js/TypeScript SDK, independent of any third-party low-code orchestrator.
- Added LLM-based category classification covering the full catalog (500+ categories) instead of a small keyword-matched subset, with a deterministic fallback, plus ticket-reference recognition with format-based disambiguation.
- Deployed as a small load-balanced cluster of workers behind a single router, and handled real operational issues (identity-sync failures, session recovery) as they came up in production.
Stack
Node.js, TypeScript, LLM-based classification, ticketing (ITSM) integration, systemd.
Outcome
A live production system handling real support traffic for more than one end client, with classification covering the full category catalog and zero dependency on the discontinued third-party platform.