prakhar2295/ai_fraud_detection_chatbot

This repository contains a Banking Fraud Detection Voice AI system written in Python. The aim, as described in the README, is to take a voice conversation, transcribe it, run it through an AI model to look for signs of banking fraud, and reply back to the caller using a synthesized voice. The whole thing is split into multiple phases, each with its own backend README, and the main page focuses on Phase 4 and Phase 7. In Phase 1, the project began as an offline pipeline: a WAV audio file was passed through speech-to-text, the text was sent to a local large language model running in Ollama, and the model produced a fraud reasoning response. Phase 2 added microphone input and near-realtime streaming. Phase 3 introduced a graph-based workflow using LangGraph, which is a library for orchestrating steps in a deterministic order, covering intent detection, fraud analysis, risk scoring, and memory. Phase 4, the main focus of this README, adds spoken replies. A new TTS layer based on Piper, an open-source text-to-speech engine, turns the AI's response into audio. There is a queue-safe playback manager, turn management to track who is speaking, and the start of interruption support so the user can talk over the bot. A new conversation coordinator wires together speech-to-text, the LangGraph workflow, and the TTS playback. Phase 7 adds long-term memory and vector-based retrieval of past fraud patterns using ChromaDB, with a fallback to an in-memory store when ChromaDB is not installed. The README documents how to install dependencies from requirements.txt, run the server with Uvicorn on port 8000, connect a WebSocket client at /ws/voice/<session_id>, and send PCM16 audio frames along with control messages like flush, stop, and ping.