LLM Customer-Support Automation
Support automation that answers from a company's own knowledge, in production.
- Role
- AI / backend engineer
- Year
- 2024
- Timeline
- ~4–6 weeks
This is an NDA-signed client project, so it isn’t linked publicly. I’ll walk you through it live on a call and answer anything about how it was built.
A live product was drowning in repetitive support questions that all had answers somewhere (in docs, past tickets and product data), but nobody could find them fast enough, so the same questions kept reaching a human.
Answers had to be grounded in the company's own content rather than hallucinated, fast enough for a chat experience, and safe to put in front of real paying customers.
Grounded every answer with retrieval-augmented generation over the company's corpus.
Instead of Prompting a base model with no grounding.
Why: RAG kept answers tied to real, current content, the difference between a helpful assistant and a confident liar.
Used hybrid semantic + keyword retrieval, with guardrails.
Instead of Pure vector search.
Why: Hybrid retrieval caught both fuzzy questions and exact product terms, and guardrails stopped it answering outside what it actually knew.
Streamed responses inside the existing product.
Instead of A bolt-on chatbot widget in the corner.
Why: Streaming natively felt fast and trustworthy, and let the system hand off cleanly to a human whenever confidence dropped.
- Ingestion + embedding pipeline with real-time indexing
- Hybrid semantic + keyword retrieval
- Contextual chat with memory & guardrails
- Streaming responses in the product UI
- Human-in-the-loop escalation on low confidence
Hemant Manglani
Ahmedabad, India
Have a product that has to work? Let’s talk.
A 20-minute call, no pitch. We work out whether this is a real problem worth solving, and if I’m not the right person, I’ll tell you.
Every message answered within one business day