Projects
Open-source code on GitHub, from Bangla speech to LLM agents. Models and datasets are listed on the home page.
Speech & voice
-
VITS fine-tuning for authentic Bangladeshi pronunciation and accent, with before/after audio.
-
End-to-end pipeline that turns YouTube audio into LJSpeech-style TTS datasets, with Bengali support.
-
Real-time front-desk voice agent built on LiveKit Agents.
-
Voice AI bot built with Pipecat, taken from local development to cloud deployment.
-
Enterprise telephony with real-time speech processing and LLM-driven conversations.
LLMs & agents
-
QLoRA fine-tuning of Qwen2.5-7B for Bangla reasoning, runnable on Colab.
-
Self-reflective, agentic retrieval-augmented generation pipeline.
-
Conversational rental assistant that collects search criteria step by step (location, bedrooms, budget, pets), with a React front end.
Vision
-
Serverless API that extracts text from images with Google Cloud Vision, built with FastAPI for Cloud Run.
-
Browser app with real-time face tracking that places 3D (GLB) glasses and hat models on the face.
ML foundations
-
Solutions to TensorTonic machine-learning implementation problems.
-
Code and experiments from PyTorch basics to advanced deep-learning models.
Software
-
Cloud-native e-commerce backend with 24 microservices and an event-driven architecture.
-
Web app that analyses coding patterns with Llama 3.1 on Cerebras to flag developer stress and burnout.
-
Blog web application built with two classmates during my B.Sc. at RUET.