Evidence Claim Pipeline (Thesis)
A system that extracts claims from scientific papers and reasons about the evidence and premises behind them, using claim verification and abductive reasoning with LLMs.
I work on large language models and natural language processing — how to make AI systems reason better, stay grounded in evidence, and work reliably across languages, speech, and real-world data. In my PhD, I want to study how LLMs can check their own reasoning against evidence, reducing hallucination, and how to make that hold up in low-resource languages like Bangla, where labelled data is scarce.
I'm currently an AI Researcher at Penta Global Ltd., working with a small research team on Bangla medical speech recognition, medical prescription understanding, and agentic test-time reinforcement learning. I went to sea first because I wanted a challenging, disciplined profession, and spent five years as a marine engineer on VLCC super tankers. In 2023 I came back to my first interest, computing: I started a B.Sc. in Computer Science & Engineering at BUBT from scratch, and since 2024 have studied alongside full-time industry work while holding a 4.00 CGPA.
- Reasoning & Claim Verification: Evidence Claim Pipeline (thesis, ongoing) · Agentic TTRL (ongoing) · LLM natural language inference (code)
- Multilingual & Low-Resource NLP: PromptGuard (BLP @ AACL 2025) · Human-Rights Dataset Extraction (AVNBD)
- Speech & Clinical AI: Bangla Dialectal Medical ASR Benchmark (ongoing) · Medical prescription understanding (ongoing)
Graduating Jan 2027 · actively seeking PhD admission (Fall 2027) in NLP & LLMs. Professors with openings — please reach out, I'd love to talk.
Recent news
- Now working mainly as a researcher on the Penta AI team: Bangla medical speech, medical prescriptions, and agentic test-time RL.
- Moved into an AI Engineer role at Penta Global Ltd., building LLM pipelines for scientific claim extraction and evidence grounding.
- Completed three DeepLearning.AI specializations: Machine Learning, Deep Learning, and NLP.
- Received the Dean's Award for Academic Achievement at BUBT for the second time (Fall 2025, after Spring 2025). Certificate →
01 Publications
-
BLP @ AACL
2025PromptGuard at BLP-2025 Task 1: A Few-Shot Classification Framework Using Majority Voting and Keyword Similarity for Bengali Hate Speech Detection
Proceedings of the BLP Workshop @ AACL-IJCNLP 2025 · 2 citations
A prompt-based, few-shot framework for Bengali hate speech detection that combines LLM classification with majority voting across repeated generations and keyword-similarity signals, targeting reliable predictions in a low-resource language setting. It reaches 67.61 micro-F1, compared with 60.75 for an n-gram baseline.
Synced from Google Scholar · Oct 30, 2026.
02 Ongoing research
Agentic Test-Time Reinforcement Learning
Applying test-time reinforcement learning (TTRL) to LLM agents, so they can keep improving on new, unlabeled tasks using their own consensus as a reward signal.
With MD Fahim, Mir Sazzat Hossain, Intesar Tahmid, Galib Mahmud Jim, Md Mahir Jawad
Bangla Dialectal Medical ASR Benchmark
A benchmark for clinically significant errors in Bangla dialectal medical speech recognition, built from synthetic doctor–patient conversations generated with LLMs and text-to-speech.
With MD Fahim, Mir Sazzat Hossain, Intesar Tahmid, Galib Mahmud Jim, Md Mahir Jawad
Medical Prescription Understanding
Reading and structuring medical prescriptions with AI, so that drug names, dosages, and instructions can be extracted reliably.
With MD Fahim, Mir Sazzat Hossain, Intesar Tahmid, Galib Mahmud Jim, Md Mahir Jawad
Human-Rights Dataset Extraction
Multi-stage LLM pipeline that turns a scanned Bangla source document into a structured, translated dataset, validated against an independent third-party dataset.
With AVNBD
03 Open source
PromptGuard
Code for our BLP-2025 shared-task paper (AACL-IJCNLP 2025 workshop): few-shot LLM classification with majority voting and keyword similarity for Bengali hate speech detection.
abductive_reasoning
Natural language inference experiments with Qwen3-8B, exploring how LLMs reason about premises and hypotheses.
ml-penta
Retrieval-augmented generation (RAG) and RoBERTa fine-tuning experiments.
transformers_architechture_3D_view
Interactive 3D visualization of how a Transformer processes a sentence, stage by stage. React + Three.js.
Emotion_Detection_hybrid_approch
This project focuses on developing a machine learning model capable of classifying text into eight distinct emotion categories: joy, sadness, fear, anger, surprise, neutral, disgust, and shame.
securegate-ai-pentest
Secure auth portal with a multi-agent security validation suite: five agents (SQLi, XSS, CSRF, lockout, headers) test whether each defense holds.
See all 27 repositories on github.com/Rakib911Hossan.
04 Education & background
B.Sc. Engg. in Computer Science & Engineering
Bangladesh University of Business and Technology (BUBT)
Pre-Sea Training in Marine Engineering
United Marine Academy
Honors & awards
- Dean's Award for Academic Achievement (twice) — Faculty of Engineering & Applied Sciences, BUBT · Spring 2025, Fall 2025
- ToolsPython · PyTorch · Hugging Face Transformers · TensorFlow · scikit-learn · Pandas · Docker
- Coursera · DeepLearning.AI (2026) Machine Learning · Deep Learning · NLP Specializations
From ship engine rooms to trustworthy AI
For five years I kept the engines of VLCC super tankers running — main engines, generators, boilers — where every fault has to be traced from its symptoms back to a root cause, and a wrong call at sea has real consequences. I bring the same habit to AI research: treat model outputs as claims that need evidence, and make systems reliable before making them impressive.
Mentors & collaborators
- Shubhashis Roy DiptaGraduate Research Assistant, UMBC
- MD FahimResearch Assistant, UT Dallas
- Mir Sazzat HossainResearch Assistant, CCDS, IUB
Let's talk research.
I'm happy to chat about LLMs, reasoning, trustworthy AI, or multilingual NLP & speech — I'm applying for PhD programs and open to research collaborations.
rakib911hossan@gmail.com