Kaushik Pranay

Customer Care RAG System

Identified repeated customer queries consuming support associates’ time; built a self-improving RAG system using Titan Embeddings V2 and Llama 3-8B via Groq that reuses verified answers instead of retraining. • Engineered a 3-stage adaptive retrieval pipeline in LangGraph (top-5 -> top-8 -> top-11 on retrieved data), widening search and lowering thresholds for low-confidence queries, with groundedness validation before responding or escalating. • Validated on a 52-query evaluation harness with per-category failure analysis: 86.67% Recall@1, 88.9% Recall@5, 80.4% MRR, 85.7% in-scope accuracy, and 90% correct out-of-scope escalation.

Naukri Automation

Pydentic