Building Plinth at Penta Global
Machine Learning Engineer · NLP Researcher
I build AI systems that turn drawings into 3D BIM models, and I research how large language and vision-language models handle low-resource languages like Bangla.
01 · About
I'm a Machine Learning Engineer at Penta Global Ltd and an M.Sc. student in Computer Science at BRAC University. At work I lead the BIM Model Automation project, which turns scaffolding drawings, PDFs and CAD files into 3D BIM models. I'm also building Plinth, an AI-powered BIM platform that brings 2D-to-3D modelling, quantity take-off and model auditing into one place.
My research is on large language and vision-language models for low-resource languages, especially Bangla and its regional dialects. My work has appeared at EMNLP 2026 (main conference, first author), Findings of ACL 2026 and the BLP 2025 workshop at IJCNLP-AACL, where our dialect benchmark won the Best Long Paper Award.
02 · Experience
Penta Global Ltd Dhaka, Bangladesh
Penta Global Ltd Dhaka, Bangladesh
Synergy Bangladesh Dhaka, Bangladesh
03 · Research
LLM and VLM evaluation for Bangla: dialects, transliteration, medical VQA and sentiment. Full list on Google Scholar.
A 6,000-utterance benchmark across five regional dialects that aligns native script, Romanized transliteration, Standard Bangla and English. Shows that transliteration consistently degrades LLMs, and that LoRA with 160 examples per dialect beats closed-source models.
Introduces BanglaMedVQA, a clinically validated medical VQA dataset, and shows that leading LVLMs fall below chance on fine-grained diagnostic categories such as condition and position.
A new annotated dataset covering four major Bangla dialects with translations and sentiment labels, used to benchmark modern LLMs under zero- and few-shot prompting.
Our BLP-2025 Task 1 system: fine-tuned BERT variants with linear probing on top, a lightweight way to improve hate speech detection in Bangla YouTube comments.
Compares traditional machine learning models with fine-tuned LLMs for sentiment classification of Bengali drama (natok) reviews, including hyper-parameter and optimizer studies.
04 · Projects
Vision-language model acuity benchmark
A psychophysics-style benchmark and LoRA fine-tuning study of the visual limits of VLMs, comparing interventions under a cost-matched evaluation protocol.
Research project
Decentralized payroll streaming protocol
A Solidity app for time-based ETH salary vesting, employee withdrawals, cancellation settlements and protocol fee accounting, with a MetaMask-connected front end.
Blockchain simulation & incentive analysis
Extended a Python blockchain simulator to study miner participation, block intervals, stale rates, transaction ordering and reward halving through configurable experiments.
Sentiment analysis as an MCP tool
Exposed TextBlob sentiment analysis through a Gradio-based MCP server and built a conversational tool client with Hugging Face smolagents.
BERT fine-tuning for review classification
Fine-tuned a pretrained BERT model with a custom classification head in PyTorch to label movie reviews as positive or negative.
Object detection in street imagery
Trained and evaluated a YOLOv8 object detection model to find shop signboards in street images.
05 · Skills
06 · Education
BRAC University, Dhaka
BRAC University, Dhaka
Cantonment English School & College, Chattogram
December 2024
7.5 overall
January 2025
302 total
07 · Contact
Open to research collaborations, ML engineering roles and conversations about applied AI. Email is the fastest way to reach me.