TMAll workRésumé PDF ↓

2026 · Client work / Customer support

Ridley Academy Support Workflow

RAG-powered email triage, automated resolution & safe human escalation.

Ridley Academy teaches over 40,000 online piano students worldwide, receiving 100–300 daily inquiries through its support@ridleyacademy.com Gmail inbox. This automated n8n workflow resolves 80–90% of routine questions using Pinecone RAG while instantly routing refunds and disputes to human agents via Slack.

Ridley Academy Support Workflow screenshot — 1 of 3
Ridley Academy Support Workflow screenshot — 2 of 3
Ridley Academy Support Workflow screenshot — 3 of 3
1 / 3

What it does

  • High-volume student support triage: Ridley Academy's global piano community of 40,000+ students generates 100–300 daily support emails through Google Workspace Gmail (support@ridleyacademy.com) and web contact forms. Routine inquiries regarding login assistance, password resets, video lesson uploads, and sheet music resources overwhelmed the team, delaying critical student responses.
  • Structured classification & safety guardrails: An automated n8n pipeline ingests incoming emails every minute, strips HTML, and passes clean text through an empty-body safety guard before querying OpenAI GPT-4o. The classifier outputs structured JSON with confidence scoring across six automated categories (login help, password reset, video upload, course access, resource requests, and technical issues) while detecting high-risk situations.
  • Strict human escalation & Slack alerts: Automated replies are strictly prevented for sensitive cases: refund requests, payment disputes, chargebacks, fraud allegations, and legal concerns unconditionally trigger escalation. A dual-factor decision gate requires both an auto_reply flag and >= 70% confidence; any failure dispatches formatted HTML summaries and immediate Slack notifications to human support agents while tagging the message in Gmail.
  • Grounded Pinecone RAG & threaded replies: For approved auto-replies, OpenAI text-embedding-3 embeds the inquiry to retrieve the top 3 verified support documents from a Pinecone vector index. The generation model drafts concise, step-by-step instructions (under 150 words) derived strictly from retrieved knowledge without hallucination, delivering answers via Gmail with intact In-Reply-To and threadId headers.