Fine-tuned Wav2vec 2.0 and Whisper models with PEFT-LoRA for Bengali Speech-to-Text (STT), conducting performance analysis on augmented datasets to identify and address model weaknesses.
Engineered a RAG-based chatbot for REVE CHAT, integrating GraphRag, LightRag, Multi-Modal Rag, Agentic Rag, and LangChain, leveraging hybrid search (BM25 + dense) for enhanced AI chatbot responses.
Optimized vits model for Bengali Text-to-Speech (TTS) to improve synthesis quality and naturalness.
Implemented model compression techniques using Olive, Quantization, and ONNX, significantly improving model loading and inference times for both online and offline services.
Developed an election commission chatbot utilizing BERT and Bi-LSTM architectures, achieving enhanced accuracy in responses.
Integrated face liveness detection for an NTMC biometric authentication project using OpenCV, CNN, and MediaPipe to prevent spoofing.
Developed robust Time Series Models and a comprehensive forecasting pipeline for precise production quantity and sales amount predictions.
Enhanced the frontend of the REVE CHAT dashboard, significantly improving user experience, usability, and interface responsiveness.