Ongoing Research
Back-propagation is a costly operation, especially in deep neural networks. In this work, we are evaluating the results (time and accuracy) of gradient prediction to replace conventional back-propagation. This work aims to reduce training duration of neural networks.

Field: Deep Learning, Distributed Learning

• SupervisorDr. Muhammad Abdullah Adnan, Professor, CSE, BUET
• My Contribution Initiation, Architectural Design

Ongoing Research
We are trying to measure the effectiveness of LLM to provide unbiased mental health support considering the diverse socio-economic conditions of Bangladesh. As the relevant data is considered highly sensitive, my current goal is to find the best-suited prompting technique to produce synthetic data and validate its viability by professionals. I convinced a reputed consultation company, LifeSpring, to collaborate with me.

Field: Ethical NLP, Mental Health

• SupervisorDr. Shubhra Kanti Karmaker (Santu), Assistant Professor, Department of CS, UCF

The Internet of Things (IoT) based real-time health monitoring system has contributed towards a brilliant human welfare both in urban and rural areas. Many of such solutions are not well applicable in developing countries like Bangladesh due to the lack of an uninterrupted communication system. In this paper, we present an IoT-based real-time health monitoring system that can measure, monitor and report people’s health condition online and offline from anywhere.

Field: Internet-of-Things, Monitoring App

• SupervisorMohammad Nuruzzaman Bhuiyan, Associate Professor, CSE, NSTU
• My Contribution Designed and implemented core idea.
• PublishedIEEE Access

Undergraduate Thesis
In our work we explored the creation of a more effective method for identifying relevant social media influencers through an ontology-based semantic search engine. Traditional keyword-based search methods often fail to identify the best influencers, so this study aims to improve the process by defining concepts, relationships, and metrics that categorize influencers based on audience location, interaction rates, and content types. We assumed the user has a primary idea of the ontology. We initiated our study again recently with a desire to address the assumptions made in our previous research by fine-tuning LLaMA2 and employing prompt engineering techniques. This approach will allow us to perform a comparative analysis between the results obtained from LLaMA2 and those from our earlier methodology. Through this comparison, we seek to assess the improvements in performance and effectiveness of these advanced techniques over the original framework.

Field: Natural Language, Semantic Search

• SupervisorDr. Muhammad Masroor Ali, Professor, CSE, BUET
• Dev Tools and Frameworks Neo4j, Django, Python, NumPy, Pandas
• Thesis Defense Slide
• Thesis Report