Contribution

Contributed the first LoRA (Low-Rank Adaptation) finetuning implementation to this large open-source TTS repository, built on Xiaomi's OmniVoice model.

Finetuned CosyVoice3, an LLM-based Text-to-Speech model, to add Bengali support that the base model did not have by default, published as an open model with a public HuggingFace Space demo.

Finetuned a LoRA adapter adapting Orpheus, an LLM-based TTS model, for Bengali speech synthesis, published on HuggingFace with a public HuggingFace Space demo.

An open-source benchmarking playground for comparing different Text-to-Speech (TTS) systems side by side, evaluating quality, voice fidelity, and performance — including Real-Time Factor (RTF) and inference latency across models.

Added benchmark scripts and a Python-to-C++ integration test for OmniVoice (a voice cloning model) in audio.cpp, a C++ audio inference engine for TTS and voice cloning. Benchmarked weight-type configurations across inference step counts on an RTX 4090, finding f16 generator weights ~1.26x faster than native, and documenting a runtime audio-tokenizer weight type that produced unusable audio.

An open-source tool for generating paired male/female speech datasets from Bengali text scripts using the ElevenLabs Text-to-Speech (TTS) API, with emotion/tone control tags and resume support — built for TTS and Automatic Speech Recognition (ASR) model training workflows.

Finetuned a LoRA adapter to study triggered brand-bias behavior in an LLM-driven webshop agent, published with a live demo.

Actively publish curated datasets and quantized model variants on HuggingFace, with several thousand+ downloads in total across published artifacts.

Contributed to the JUnit testing framework - one of the most widely used testing frameworks in the Java ecosystem.

Developed and published as open-source library: SemanticAIMapper - a deterministic-first semantic mapping mechanism using structured LLM planning and validation. The library helps align multiple structured representations across storage, interfaces, messaging, analytics, and application layers using bounded LLM planning.

Developed and published as open-source library: THAI - a latency-agnostic framework for intelligent malicious HTTP request detection using Large Language Models. The framework provides a two-phase mechanism for WAF/IDS systems with multi-request batching that reduces LLM overhead by 90%+.

Contributed ground-truth test suites and assisted with evaluation and validation for AutoTestGen, a multi-agent LLM-based pipeline for end-to-end web test generation, an undergraduate thesis supervised by Sukarna Barua, Associate Professor at BUET CSE.

Architectural framework for intelligent malicious HTTP request detection using LLMs with two-phase mechanism (deterministic pre-filtering + async LLM confirmation). Multi-request batching reduces LLM overhead by 90%+ while eliminating 98% of false positives.

Topics: Cybersecurity, LLM, HTTP Security, Machine Learning

Object-oriented design for CRUD scaffolding with reusable, type-parameterized base components. Features template-based Controller-Service-DAO pattern with navigable retrieval.

Topics: Software Architecture, CRUD, Scaffolding, Object-Oriented Design

Full implementation of Congestion-Control AODV (CC-AODV) in NS-3 simulator with performance evaluation. CC-AODV improves throughput and packet delivery ratio in denser networks through congestion-aware route discovery.

Topics: MANET, AODV, Congestion Control, Network Simulation

Deterministic-first semantic mapping mechanism using bounded LLM planning. Resolves structured data alignment across storage, interfaces, messaging, and analytics layers with validated output schemas.

Topics: LLM, Semantic Mapping, AI Planning, Structured Output

Analysis of which occupations in developing and lower-income economies are most resistant to full automation by AI.

Topics: AI, Labor Economics, Automation, Developing Economies

Completed the Deep Learning Specialization on Coursera, covering neural networks, hyperparameter tuning, convolutional networks, sequence models, and practical ML project structuring. The specialization provides hands-on experience with TensorFlow and modern deep learning architectures.

Link: Course Certificate

Singlehandedly organized the photography exhibition at BUET CSE Fest 2022 — managed participant submissions, judged and scored entries, and awarded prizes to winners.