Customer operations AI
AI Customer Service Assistant
A RAG-powered chatbot that helps customer service reps answer complex product questions and perform financial calculations in real time.
I built this as an AI-powered knowledge assistant for customer service representatives handling complex insurance product inquiries. Instead of searching through dense product specifications, reps could ask a question in plain language and receive an answer grounded in the relevant source material—along with the calculation or next step when the question called for one.
How It Works
The system combines retrieval-augmented generation with a smart question routing layer that classifies every question into one of 18 categories — then handles each differently:
- Product questions trigger semantic search against a vector database of product specifications, injecting the most relevant context into the prompt for accurate, grounded answers
- Financial calculations launch an interactive calculator for withdrawal amounts, premium bonuses, and disbursements — with formula visualization and step-by-step walkthroughs
- Simplification requests take a previous answer and rewrite it in plain language suitable for explaining to customers
- Operational questions route to the correct internal system — knowledge bases, document management, help desks — with specific instructions on where to go
Architecture
The question routing layer classifies incoming questions and dispatches them to the appropriate handler. For product questions, the system performs semantic search against a vector database of indexed product specifications, retrieves the top matching passages, and injects them as context into the LLM prompt. Temperature and sampling parameters are tuned for consistency and reproducibility. An activity logging pipeline captures every interaction — question type, prompt template, response content, and feedback scores — for continuous evaluation and improvement.
Built-In Quality Feedback
Every response includes a feedback widget where reps can rate answer quality and leave comments. Combined with comprehensive activity logging (conversation IDs, timestamps, turn counts, question classifications), this created a continuous feedback loop for measuring and improving answer accuracy over time.
The Result
The result was a practical layer between complex product documentation and the service conversation: less time spent hunting for source material, more consistent answers, and a clearer way to explain complex financial products to customers.
