All projects
shipped2024

Support Automation

LLM-powered support routing that actually works.


An intelligent support triage system that classifies, routes, and partially resolves incoming tickets — without the hallucination risk that killed previous attempts.

The Problem

High-volume support queues with inconsistent routing led to long resolution times and poor CSAT scores. Previous AI attempts hallucinated responses and eroded user trust.

The Approach

Built a constrained generation system that limited model outputs to verified knowledge base content. Added a confidence threshold gating layer — below threshold, it routes to human; above, it responds autonomously.

The Outcome

CSAT scores improved 40%+. Autonomous resolution rate reached 58% of tickets. Zero hallucination incidents in production after the constrained generation approach.


Technologies

TypeScriptClaude APINext.jsPineconeVercel