Publications: AdaLLM, HornNet & More

Large Language Models

Investigating emergent capabilities, fine-tuning strategies, and alignment in foundation models.

Artificial General Intelligence

Exploring architectures for general reasoning and autonomous agent systems.

Deep Learning Theory

Neural architecture optimization and efficient training methodologies.

Computer Vision

Medical imaging, real-time detection, and multi-modal perception systems.

Multi-Agent Reinforcement Learning

Coordination, credit assignment, and stable policies when several agents learn under partial observability.

Publications

Peer-reviewed conference papers

01

2026 iCACCESS 2026 Accepted

AdaLLM: A Subject-Adaptive RAG Framework Designed for the Question Answering Process of NCTB’s Class-8 Science & BGS Subjects

Authors: Samia Rahman, Raiyan Islam, Redwan Ahamad Samir, Md Kaviul Hossain, Md. Omur Faruk, Raihan Kibria

2nd International Conference on Advances in Computing, Communication, Electrical, and Smart Systems (iCACCESS 2026), October 2–3, 2026

Abstract

Bangladesh has over 160 million Bengali speakers where the students study from the textbooks offered by the National Curriculum Textbook Board of Bangladesh (NCTB). General-purpose AI chatbots give unreliable answers to these students because they are not trained on NCTB content and hallucinate confidently wrong facts, which poses a serious risk in education. This paper presents AdaLLM, a subject-adaptive Retrieval-Augmented Generation (RAG) system for NCTB Class 8’s Bangladesh and Global Studies (BGS) and Science subjects. The proposed system uses an eight-layer pipeline (L0–L7) in which L0 rewrites informal or Banglish student queries into academic Bengali, L1 detects the school subject (95% accuracy), L2 performs hybrid dense+sparse retrieval with subject-adaptive weighting (α = 0.60 for BGS, 0.40 for Science), L3 re-ranks chunks by curriculum relevance, L4 validates context quality, L5 triggers a cross-subject fallback on failure, L6 generates a Bengali answer using Gemma 3 4B, and L7 applies 5-axis confidence scoring. Textbooks were digitized via a multi-step OCR pipeline achieving CER 4.65% and WER 8.84%. We collected a human-annotated evaluation dataset of 100 Q&A pairs with 1,000 student-written answers and 1,000 informal query variants verified by subject-matter teachers. Our full system achieves ROUGE-L score of 0.1654 and BERTScore F1 of 0.7120 on informal student queries, outperforming RAG without rewriting (ROUGE-L 0.1088) and direct LLM generation without RAG (ROUGE-L 0.1225). On formal academic queries the system achieves a Hit Rate of 0.71, MRR of 0.5603, and Faithfulness of 0.7859 with a hallucination rate of 0.0.

Keywords

Retrieval-Augmented Generation Bengali NLP Educational AI Low-Resource Languages Query Rewriting Hybrid Retrieval Hallucination Mitigation NCTB Curriculum

02

2026 IEEE QPAIN

HornNet: A Two-Tier Deep Learning Framework for Real-Time Vehicle Classification from Acoustic Horn Signatures in Urban Environments

Authors: Fakir Mashuque Alamgir, Redwan Ahamad Samir, Sayed Easin Arafat, Samia Rahman, Abidur Rahman

2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN)

Abstract

This paper presents HornNet, a two-tier deep learning framework for real-time vehicle classification based on acoustic horn signatures in urban environments. The proposed architecture addresses limitations of vision-based traffic monitoring systems by leveraging a hierarchical structure: a lightweight edge detector for initial sound event detection and a deep classifier for vehicle type identification. Our approach enables privacy-preserving, low-cost traffic monitoring suitable for smart city deployment.

Keywords

Audio Classification Smart Cities Edge AI Acoustic Monitoring Deep Learning Traffic Analysis

DOI PDF Code

03

2025 BMI Conference

Early-Stage Coronary Artery Disease Prediction Using Coronary Angiogram for Stenosis Detection in Low-Resource Settings

Authors: Fazlay Rabby, Musfiqur Rahman, Rajdeep, Hridoy Hossain, Redwan Ahamad Samir, Sfarhaan

International Conference on Big Data, IoT and Machine Learning (BMI 2025)

Abstract

Coronary Artery Disease (CAD) is a leading cause of death worldwide, emphasizing the need for early and accurate detection. This study presents a deep learning framework for automated stenosis detection from coronary angiograms, specifically designed for low-resource clinical settings. Our approach leverages convolutional neural networks with attention mechanisms to identify arterial blockages with high accuracy, enabling early intervention and improved patient outcomes in resource-constrained environments.

Keywords

Coronary Artery Disease Medical Imaging Deep Learning Computer Vision Healthcare AI Low-Resource Settings

PDF Code

04

2025 BMI Conference

Scalable sEMG-Based User-Independent Deep Learning Framework for Assistive Wheelchair Control

Authors: Fazlay Rabby, Rajdeep, Musfiqur Rahman, Hridoy Hossain, Redwan Ahamad Samir

International Conference on Big Data, IoT and Machine Learning (BMI 2025)

Abstract

Surface electromyography (sEMG) based wheelchair control improves the quality of life for people with severe disabilities. This study presents a low-cost, real-time, user-independent deep learning framework for assistive wheelchair navigation. By processing multi-channel sEMG signals through a novel neural architecture, our system achieves robust gesture recognition across different users without individual calibration, significantly reducing deployment barriers for assistive technology.

Keywords

sEMG Signal Processing Assistive Technology IoT Deep Learning Accessibility

PDF Code

Current Research

Ongoing investigations and future directions

Active

Multi-Agent Reinforcement Learning

Studying how several learning agents coordinate, compete, and share information inside one environment. The work looks at cooperative decision making when each agent sees only part of the world, and at what it takes for multi-agent policies to stay stable and scale to real-world coordination problems.

Research Objectives:

Design cooperative and competitive multi-agent learning setups
Study credit assignment and coordination between agents
Handle partial observability and non-stationarity when many agents learn at the same time
Evaluate learned policies for stability, scalability, and fairness

Multi-Agent RL Reinforcement Learning Multi-Agent Systems Decision Making

Active

Agentic RAG

Extending retrieval-augmented generation from a fixed retrieve-then-answer pipeline to agent-driven workflows. An LLM agent plans the steps, decides when and what to retrieve, calls tools, checks the evidence, and iterates before it answers. This builds on my earlier work on subject-adaptive retrieval for educational question answering.

Research Objectives:

Plan multi-step retrieval and decide when retrieval is needed
Route and rewrite queries across multiple knowledge sources and tools
Verify retrieved evidence and answers to reduce hallucinations
Evaluate retrieval quality, faithfulness, and cost of agentic pipelines

Agentic RAG LLM Agents RAG NLP Tool Use

Interested in Collaboration?

I am always open to research collaborations, particularly in areas related to large language models, AGI safety, and applied deep learning for social good.

Get In Touch