Agentic Voice Bot for BFSI(using RAG)
Developed an on-premises Agentic Voice Bot for BFSI using Pipecat, integrating Silero VAD, STT, LLM, and TTS components with RAG using FAISS for knowledge retrieval and an agentic workflow for intelligent, context-aware voice interactions.
Bank Statement Aggregator with Fraud Detection for BFSI (2025–26)
Developed client-specific fraud-detection logic optimized to O(n) time and space complexity, enabling concurrent processing of thousands of client logins and lakhs of transaction rows, and deployed it on-premises via FastAPI with SQL/ODBC connections, webhook-based data exchange, and AWS service integrations.
Fine-Tuned ViT-Based Model for Selfie Clear/Blur Detection (2025–26)
Engineered a multi-model image classification system for clear/blur selfie detection by fine tuning two open-source ViT models across three distinct approaches, integrating them with a base model and an LLM into a unified four-model architecture that achieved 99.98% accuracy. Owned end-to-end development, including iterative hyperparameter tuning, and deployed the pipeline on AWS (Lambda, S3, API Gateway, SageMaker, Bedrock, ECR) using Docker.