Embedded Files
LLMs for Churn Prediction with Reasoning to Enhance Trust
A Retrieval-Augmented Generation (RAG) enhanced Large Language Model offers a highly effective and explainable approach to customer churn prediction by combining the retrieval of relevant customer data with advanced reasoning capabilities. The system first queries a vector-based knowledge base containing customer profiles, usage history, support interactions, billing records, and other unstructured data to fetch the most pertinent context. This retrieved information is then fed into the LLM, which performs step-by-step reasoning to assess churn risk, identify key contributing factors (such as declining engagement, negative sentiment in tickets, or upcoming contract expiry), and generate a clear natural language explanation along with a churn probability rating. Unlike traditional black-box models, this method delivers transparent, actionable insights and personalized retention recommendations while easily adapting to new business policies or data sources without requiring full model retraining, making it particularly valuable for industries like telecom, SaaS, banking, and e-commerce.
Agentic AI System for Customer Support
This paper presents an Agentic AI-based Customer Support System that integrates Retrieval-Augmented Generation (RAG) with a multi-stage workflow to provide accurate and context-aware responses to customer inquiries. The system employs sentence-transformer embeddings and a FAISS vector database to retrieve relevant knowledge from a customer support knowledge base. A local Llama 3.1 model orchestrated through LangGraph performs three sequential tasks: customer query analysis, response generation, and follow-up question recommendation. The framework maintains conversation history to support contextual understanding and personalized interactions. Implemented using Flask, Ollama, and LangGraph, the proposed architecture minimizes hallucinations by restricting responses to retrieved knowledge and generates actionable answers with suggested follow-up questions. Experimental evaluation demonstrates the effectiveness of the system in handling common customer service requests, including account management, shipping, returns, and payment issues. The proposed approach offers a secure, scalable, and cost-effective solution for intelligent customer support automation.
This project is for academic advising assistant built specifically for the Colleges/University. It combines four core technologies: Flask for the web server, LangGraph for orchestrating a multi-step AI agent, FAISS for semantic search, and Ollama running the llama3.1:8b model locally (so there are no API costs).
The system works as an agentic pipeline with three sequential nodes. The first node analyzes each student's question to identify its topic category and underlying intent, then uses that analysis to search a knowledge base for relevant facts via vector embeddings and FAISS similarity search. The second node takes the retrieved context and generates a structured, encouraging response: a direct answer, 2-4 actionable steps, and an uplifting closing line, while being instructed to only use facts from the knowledge base rather than inventing information. The third node generates three relevant follow-up questions as a JSON array, helping students continue the conversation naturally.
The knowledge base currently contains just one paragraph covering UWF's founding year, location in Pensacola, size, and mascot (Argie the Argonaut). This is quite limited — to make the advisor genuinely useful, it would need substantially more UWF-specific content covering academic programs, advising procedures, financial aid, housing, and student services.
Page updated
Google Sites
Report abuse