Selected work

Projects, annotated.

Things I have built, each useful, and each a small experiment in turning research into product.

Research artifacts

BanglaMedVQA · Findings of ACL 2026

Clinically validated bilingual medical VQA benchmark: 7,000 image–question pairs across five question types, with LoRA fine-tuning code.

Dataset ↗Code ↗Project page ↗

LP-FT-LoRA · BLP Workshop, IJCNLP-AACL 2025

Three-stage PEFT framework: align the classifier head with a linear probe, train LoRA adapters, then refine both.

Code ↗

Applied AI

AI · pipelines

№ 01

Stock-Agent

An automated intelligence pipeline that scrapes financial news, uses local AI models like FinBERT and Phi-3 for sentiment analysis and summarization, and generates daily professional PDF insights.

Developed an automated intelligence pipeline scraping financial news using Selenium and BeautifulSoup
Engineered a local-first AI analysis layer using FinBERT for sentiment and Phi-3 for concise summarization
Implemented NER and fuzzy matching scripts to extract and map company entities to DSE ticker symbols

Python · Transformers · Ollama · Selenium · WeasyPrint

Earlier software projects

Web · full-stack

№ 02

MediKey

A full-stack healthcare web application that simplifies the process of finding doctors and booking appointments. Integrated an AWS-powered intelligent chatbot that analyzes symptoms and recommends suitable specialists.

Spring Boot · React.js · Amazon Web Services · Java

Code →

Tool · NLP

№ 03

JobGenie

A smart web application that auto-generates a professionally designed CV from user-provided details and scrapes real-time job listings tailored to the user's profile and qualifications.

FastAPI · React.js · MongoDB · Python

Code →

Mobile · social

№ 04

OyeAmigo

A Kotlin-based Android social media application that helps users discover and connect with others who share similar preferences. Built with a focus on performance and clean code using Kotlin's modern language features.

Kotlin · Android SDK · Firebase

Code →

Each project answers a small, concrete question. The interesting ones leave behind a method, not just a demo.