AI customer support assistant — FAQ matching, sentiment-aware escalation, and LLM-generated responses. Built with Amazon Bedrock (Claude) and FastAPI.
Portfolio: rajeshdevandla.github.io
View the full project overview, architecture, and demo walkthrough on the portfolio site.
To run locally, follow the Quick Start below.
Most customer support workloads are repetitive: the same 20 questions account for 80% of tickets. SupportGPT handles those automatically via a FAQ layer, uses sentiment analysis to detect frustrated customers before they escalate themselves, and falls back to Claude for everything else. Human agents only see the hard cases.
Customer: "How do I return an item?"
SupportGPT: [FAQ match] "Returns are accepted within 30 days with original receipt..."
handled_by: faq | sentiment_score: 0.72 | escalate: false
Customer: "I've been waiting 3 weeks and NOBODY is helping me this is unacceptable"
SupportGPT: [Escalating to human agent — frustrated customer detected]
handled_by: escalation | sentiment_score: 0.18 | escalate: true
Customer: "Can you explain your enterprise pricing options?"
SupportGPT: [Bedrock] "Our enterprise plans are customized based on team size and usage..."
handled_by: llm | sentiment_score: 0.65 | escalate: false
Ticket summary (auto-generated at end of session):
Session: 3 messages | Duration: 4 minutes
Topics: returns policy, shipping delay, enterprise pricing
Escalated: Yes (message 2 — high frustration detected)
Recommended action: Follow up on shipping delay