Email Triage Agent Comparing Rule-Based n8n Workflows with an LLM Agent

Should a helpdesk triage its e-mails with keyword rules in a workflow tool, or with an LLM agent? This project builds both on a standard laptop and measures them on the same 150 labelled e-mails: an n8n workflow (Webhook -> Code node with keyword rules learned from 11,736 TRAIN e-mails -> Respond to Webhook), a tool-using agent on qwen2.5:1.5b in local Ollama (plan tool calls -> similar-ticket search and rule check -> JSON-schema decision -> guards), a single LLM call, a hybrid router and a TF-IDF + logistic regression classifier. Each gives intent, priority, queue and a reply suggestion for a person.

NOTE: Teaching study on synthetic helpdesk e-mails (no personal data). Labels and replies are suggestions for a person who reads the e-mail; nothing is sent to a customer automatically. The dataset's labels are synthetic and noisy.

Category: Department of Computer Science and Engineering / Agent Development

Agent Development Project AG25 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.