Tanmoy Shome
Artificial Intelligence EngineerResearch AssistantHey Hi! This is Tanmoy here, I am currently an Artificial Intelligence Engineer at VISIE Limited where I work on integrating AI/ML models and intelligent systems into production software. I completed my undergraduate at Daffodil International University, where my thesis was supervised by Prof. Shohel Arman. During my undergraduate years, I also worked at Bengali.AI on Bengali natural language processing research, under the supervision of Prof. Swakkhar Shatabda, Prof. Farig Yousuf Sadeque, Ahmed Imtiaz Humayun, Tahsin Reasat and Asif Shahriyar Sushmit. Outside of research and engineering, I am always excited to explore new hiking and trekking destinations. Climbing hills feel like an achievement to me, and each taller peak becomes the next milestone. This year, I have been focusing mostly on physical activities and learning classical music.
Experience
April 2024 - Present
AI Engineer - VISIE Limited
Optimized AI-agent and RAG pipelines, cutting average latency from 2.1s to 850ms. Built conversational agents for customer, product, payment, and retention-risk analysis, reducing manual reporting turnaround by 60% across 20+ business KPIs. Developed credit-risk scoring pipelines for repayment and default prediction, handling class imbalance and explainable risk scores.
July 2025 - October 2025
Research Assistant - BRAC University
Investigated how foundation models understand and generate culturally grounded satire in low-resource Bangla contexts. Built a curated dataset from Unmad magazine's archive — annotated for humor type, cultural referents, and satirical target. Designed and developed Unmad Bot, an end-to-end conversational system combining retrieval over the annotated archive with an instruction-tuned multilingual model, ensuring generated humor stayed grounded in real cultural precedent rather than stereotype.
March 2023 - March 2025
Research Assistant - Bengali.AI
Benchmarked and fine-tuned document layout models (Detectron2, YOLO, SwinDocSegmenter) on a 3K Bangla newspaper dataset, and ASR models (Wav2Vec2, Whisper) on 100+ hours of regional Bangla speech with dialect variation. Contributed to a cross-domain 10K-hour Bengali speech corpus for recognition, diarization, and synthesis (RSGI-funded), and coordinated a newspaper-domain Bengali layout analysis system (IAR-funded).
Research
Digitizing an Under-Resourced Script: BaNLAD, a Layout Analysis Dataset for Bengali Newspapers
Tanmoy Shome, Radeen Mostafa, Mirza Nihal Baig , +4 Authors, Asif Sushmit, Farig Sadeque, Swakkhar Shatabda
SUBMITTED
TL;DR: We introduce BaNLAD, a dataset of 2,726 annotated Bengali newspaper images spanning historical newspapers, standard layouts, and degraded conditions including sunlight, crumpling, staining, and brown marking. With over 644,000 human-annotated polygonal segments covering text boxes, paragraphs, tables, images, and news items, BaNLAD provides the first large-scale benchmark for document layout analysis and OCR in this under-resourced language.
From Archive to Agent: Generative AI and the Mediation of Satirical Voice in Unmad Bot
Nusrat Jahan Mim, Md Ishmam Tasin, Md. Rezuwan Hassan, Tanmoy Shome, Zumaina Islam, Syed Ishtiaque Ahmed, Farida Chowdhury, S M Taiabul Haque,
SUBMITTED
TL;DR: Through the design and deployment of Unmad Bot, a conversational artifact grounded in the archive of Bangladesh's longest-running satirical magazine, we examine how generative mediation makes rhetorical distance negotiable and redistributes agency over satirical stance. We argue that generative AI should be studied not only through the content it produces, but through the communicative positions and forms of authority it constitutes within culturally situated practices of critique.
Are ASR foundation models generalized enough to capture features of regional dialects for low-resource languages?
Tawsif Tashwar Dipto, Azmol Hossain, Rubayet Sabbir Faruque, Md. Rezuwan Hassan, Kanij Fatema, Tanmoy Shome, +7 Authors, Farig Sadeque, Tahsin Reasat
IJCNLP-AACL 2025
TL;DR: We introduce Ben-10, a 78-hour annotated Bengali dialectal speech-to-text corpus, and show that speech foundation models degrade sharply on regional dialects in both zero-shot and fine-tuned settings. Dialect-specific training mitigates this gap, and the dataset doubles as an out-of-distribution benchmark for low-resource ASR.
Get In Touch
Dhanmondi, Dhaka
[firstname].[lastname].6174 [at] gmail [dot] com
© 2026 All Rights Reserved.