ML & NLP Researcher · Low-Resource Language Technology · Prospective MS/PhD Student
I build machine learning for languages and sounds that most models never see — Sylheti in both of its scripts, and the swara structure of Indian classical flute.
Most of my work starts from the same problem: the language or the signal I care about has no dataset, so I build one first. The modelling follows from the data — byte-level translation for a script no pretrained model has seen, parameter-efficient fine-tuning for Bangla classification, attention-guided segmentation for melody.
B.Sc. in Computer Science and Engineering, Sylhet Engineering College (affiliated with SUST), 2022–2026. Reviewer, ICCA 2026.
| Repository | What it does |
|---|---|
| Syltrans · live | Bidirectional Bengali ⇄ Syloti Nagri transliterator — rule-based Flask engine, converts as you type |
| multilingual-translator-web | 45+ languages with NLLB-200-distilled-600M and automatic language detection; 84.8% detection accuracy across 46 languages |
| summer-and-winter-translator | Unpaired image-to-image translation with CycleGAN in PyTorch, wrapped in Gradio |
Sylheti and Indian classical flute have almost no machine-readable resources, so each of these was collected from scratch.
| Dataset | Size |
|---|---|
| Bangla–Syloti Nagri Parallel Corpus | 21,181 aligned entries, dual script |
| SylEmotions | 7,051 sentences, 7 emotion classes, dual script |
| SwaraSet | 748 swara recordings + 1,000 synthetic melodies |
Details on the datasets page.
Two peer-reviewed IEEE conference papers, four manuscripts under review, one project in progress.
ICCIT 2025 — Beyond Sentiment: The Influence of Game Pricing on Player Feedback in Steam Reviews · DOI: Read Paper
IEEE QPAIN 2026 — Hierarchical Context-Aware Network (HC-Net): A Multi-Task Learning Approach for Robust MRI Disease Diagnosis · DOI: Read Paper
Under review: swara boundary detection in Indian classical flute · SylEmotions · equity-aware parliamentary speech attribution · Bangla misogyny detection. Full list on the publications page.
Models BanglaBERT · XLM-RoBERTa · Qwen2.5 · ByT5 · DistilBERT · NLLB-200 | Methods LoRA / PEFT · SHAP · CNN · CRNN · CycleGAN | Audio MFCC · Log-Mel · CQT · pitch & RMS · DTW
Home · Background · Publications · Datasets · Projects · Achievements
Open to MS/PhD opportunities in NLP and machine learning.