Hi! I’m Tasmiah Tahsin Mayeesha, a Ph.D. researcher in Information Science (Data Science) at the University of North Texas, where I started in Fall 2025. My current research interests are broadly centered around AI safety, evaluation, multilingual NLP, multimodal models, and the robustness of foundation models.
Before starting my Ph.D., I worked as a researcher in Bangladesh across NLP, HCI, and AI policy. I worked with the EBLICT–Dream71 collaboration on Jiggasha AI, to curate the a Bangladesh government-funded public Bangla virtual assistant developed to make public-administration information more accessible to citizens.
Concurrently, I also worked as an HCI researcher at the Design, Inclusion & Access Lab (DIAL) at North South University, working with Dr. Nova Ahmed on projects spanning participatory design, education, and responsible AI. My work included studying the experiences of women in computing through fictional inquiry and co-design, as well as examining AI ethics and governance from the perspectives of South Asian and Bangladeshi communities. This work resulted in publications at venues including ACM ToCHI, ICTD, ACM COMPASS, and UbiComp.
Alongside my HCI work, I was a senior research assistant with Dr. M. Rashedur Rahman on Bengali NLP and Natural Language Generation, where I worked on answer-aware question generation, Bengali question generation, and multimodal question generation using images. I also completed a predoctoral fellowship at the Fatima Institute for Global AI Research, where we evaluated formality bias in multilingual models across multiple languages. This work resulted in a publication in EMLNLP.
My earlier work in NLP began during my undergraduate studies in Computer Science and Engineering at North South University, where I graduated in 2020. My thesis focused on Bengali question answering using multilingual BERT and synthetic data. During my undergraduate years, I also participated in Google Summer of Code with the TensorFlow Hub team and worked with the Berkman Klein Center for Internet & Society.
These days, I’m particularly interested in understanding how and why foundation models fail—especially when reasoning is affected by adversarial inputs, language, context, or modality—and in building better ways to evaluate those failures.
Outside research, I enjoy anime, manga, books, and taking care of my cats. 🐈⬛
- 🔭 Papers: Google Scholar
- 👨💻 Website: tahsin-mayeesha.github.io
- 📝 Writing: Substack
- 📫 Contact: tasmiahtahsinmayeesha@gmail.com
| Resource | Type | What's inside |
|---|---|---|
| Bengali-SQuAD | Dataset | Bengali SQuAD dataset for question answering and question generation research |
| MisdirectVQA-200 | Dataset | 200-question diagnostic benchmark for evaluating text-only adversarial misdirection in vision-language models |
| vqa_bn | Dataset | Bengali VQG dataset based on VQA v2.0, with 220K translated training QA pairs and 150K translated validation QA pairs |
| tofu_bn | Dataset | LLM-unlearning dataset - English–Bangla machine-translated adaptation of TOFU: A Task of Fictitious Unlearning for LLMs with full, forget and retain splits. |
| t5-end2end-questions-generation | Model | T5-based end-to-end question generation model |
| squad-bn-mt5-base2 | Model | mT5-based Bengali question generation model trained on Bengali SQuAD |
| wav2vec2-bn-300m | Model | 300M-parameter Wav2Vec2 model for Bengali automatic speech recognition |
| Bangla Question Generation | Space | Interactive demo for Bengali question generation |
