Tasnimul Hossain Tomal

AI Engineer | Researcher
Penta Global Limited

tomal.tasnimul@gmail.com
github.com/tomal66
linkedin.com/in/tomal66

I am an AI engineer with a research foundation in distributed learning security and applied NLP. My research journey began with the investigation and mitigation of data-driven attacks on distributed learning paradigms. My current work is centered on LLMs with a focus on knowledge distillation techniques, fine-tuning strategies, and efficiency improvements for domain-specific applications.

Currently, I am seeking graduate opportunities to contribute to advancing both the theoretical foundations of distributed AI security and the practical implementations that make AI systems trustworthy, scalable, and impactful. I am also interested in collaborating on research that integrates robustness, efficiency, and adaptability into next-generation NLP and LLM systems.

Education

2020—2025

University of Dhaka

B.Sc. in Computer Science and Engineering (GPA: 3.57/4.0)

Research Experience

Undergraduate Thesis: A Novel Framework to Mitigate Data Poisoning Attacks in SplitFed Learning

Prof. Dr. Md. Abdur RazzaqueSupervisorJan 2024–Present

Developed Centinel, a novel framework for SplitFed learning that mitigates data poisoning attacks by employing a centroid-shift-based anomaly detection mechanism. Demonstrated the framework's effectiveness in maintaining model accuracy and integrity against malicious clients with different attack scenarios.

Distributed Machine LearningFederated & Split LearningSecurity in AI

[Under Development, manuscript in preparation]Publications

Jailbreaking & Hallucination Mitigation in (Vision) Language Models

Md. Fahim & Farhad Alam BhuiyanSupervisorOctober 2025–Present

Ongoing research and experimental study on mitigating (object & perceptual) hallucinations in LVLMs and stress-testing defenses under adaptive jailbreak attacks.

Multimodal LLM SafetyHallucination MitigationAdversarial ML

[Under Development]Publications

Improvised LoRA Finetuning for Domain-Specific Classification tasks in LLMs

Md. Fahim & Farhad Alam BhuiyanSupervisorAugust 2025–October 2025

Investigated challenges of traditional LoRA finetuning for classification tasks in unknown domains where LLMs lack prior knowledge. Proposed an improvised method by introducing Further Pretraining of LoRA Adapters (FPT-Lora), which first pretrains adapters on a synthetic corpus built from classification dataset features, followed by task-specific finetuning. Demonstrated improved adaptability and performance in domain-specific classification.

Large Language ModelsParameter-Efficient FinetuningDomain AdaptationNLP

Submitted to LREC 2026Publications

Publications

AAAI Bridge AIMedHealth 2026

How Good LLMs Are at Answering Bangla Medical Visual Questions? Dataset and Benchmarking

Rafid Ahmed, Intesar Tahmid, Mir Sazzat Hossain, Tasnimul Hossain Tomal, Md Fahim, Md Farhad Alam

Paper

BanglaMedVQA, the first clinically validated Bangla medical visual question answering benchmark, and shows that current LLM/LVLMs perform poorly on it, revealing major gaps in fine-grained medical reasoning for low-resource languages.

BLP 2025

LP-FT-LORA : A Three-Stage PEFT Framework for Efficient Domain Adaptation in Bangla NLP Tasks

Tasnimul Hossain Tomal, Anam Borhan Uddin, Intesar Tahmid, Mir Sazzat Hossain, Md Fahim, Md Farhad Alam

Paper

A three-stage framework that decouples head alignment from representation learning to address LoRA training instability.

BLP 2025

Benchmarking Large Language Models on Bangla Dialect Translation and Dialectal Sentiment Analysis

Md Mahir Jawad, Rafid Ahmed, Ishita Sur Apan, Tasnimul Hossain Tomal, Fabiha Haider, Mir Sazzat Hossain, Md Farhad Alam

Paper

A novel 600-sample Bangla dialect dataset (Chattogram, Barishal, Sylhet, Noakhali) with translations and sentiments, showing transliteration boosts on SOTA LLMs.

Submitted Manuscripts

Decoupling Representation for Fine-tuning with Linear Probing and LoRA on Robust Downstream Adaptation

Tasnimul Hossain Tomal, Intesar Tahmid, Rafid Ahmed, Mir Sazzat Hossain, Md Fahim, Md Farhad Alam

PaperSubmitted

Evaluating Vision-Language Models on Bangla Medical Visual Question Answering: Dataset and Benchmarking

Rafid Ahmed, Mir Sazzat Hossain, Intesar Tahmid, Tasnimul Hossain Tomal, Md Fahim, Md Farhad Alam

PaperSubmitted

Work Experience

Apr 2025 – Present

AI Engineer — Penta Global Limited

Building forecasting pipelines, fine‑tuning LLMs, and deploying Retrieval‑Augmented Generation services for enterprise clients.

Jan 2025 – Apr 2025

Software Engineering Intern — Data Elysium Software Inc.

Developed and maintained Next.js micro‑frontends and helped ship Affortable.AI, a low‑cost AI platform.

Sept 2023 – Present

Research Assistant — Green Networking Research Group, CSEDU

Investigating trust‑preserving collaborative and distributed learning frameworks (Split/Federated Learning).

Portfolio

AffortableAI

Next.jsVercelPostgreSQL

Project

A low‑cost AI platform with pay‑as‑you‑go pricing, enabling users to chat with LLMs securely and privately.

JobGenie

ReactFastAPIMongoDB

AI‑assisted CV builder and job‑search site that crawls listings and recommends opportunities tailored to a user’s profile.

MediKey

ReactSpringBootMySQLAWS

PaaS web application enabling paperless access to medical data with secure EHR storage and sharing between patients and doctors.

2mins2goods

ReactSpringBootMySQL

Hyper‑local marketplace supporting local entrepreneurs; integrates geolocation APIs to buy and sell goods within the neighbourhood.