Graduate Research Student
University of Waterloo

I am working with Professor Daniel G. Brown on the ability of LLMs to respond appropriately and consistently to sensitive topics with prompt variations. We analyzed over 30 models, both open and closed source, compared their performance in various areas like task understanding and response consistency. Surprisingly, our findings indicate that most models, even some large open-source models have a difficult time answering simple Yes/No questions and can barely understand the task at hand. They are prone to changing their responses with slight variations in prompt wording and have different responses in different settings. These findings warn us against using LLMs for Q/A without proper planning and testing given their limited instruction instruction-following and task-understandig capacity.

• Analyzed the capabilities and limitations of LLMs (Large Language Models) in assessing and updating sensitive statements (e.g. stereotypes, conspiracies, etc). Used HuggingFace, Python, and Tableau.
• Identified appropriate LLMs by creating a dashboard based on fact-checking, bias detection and instruction-following abilities across 37 open and closed source models. Link.
• Formulated story worlds and seamless entertaining stories by tweaking the information retention and critical assessment ability of LLMs based on aforementioned research findings.

Related Publications:

• A Study on Large Language Models' Limitations in Multiple-Choice Question Answering. (Arxiv)
• Reliability Check: An Analysis of GPT-3’s Response to Sensitive Topics and Prompt Wording. (TrustNLP 2023. ACL)

Research Data Scientist (NLP)
Wikimedia Foundation

I worked with the Research Team as a Research Data Scientist (NLP) to develop Copyediting as a structured task. Meta page, Report, Code.

• While there are ongoing efforts to automatically detect "commonly misspelled" words in English Wikipedia, most other languages are left behind. We increased the standard of articles through automated copy-editing by curating and detecting commonly misspelled words in 100+ languages.
• Used Python and PySpark to extract and analyze Wiktionary data and to detect misspellings in Wikipedia.

Currently, I am working on improving the link-recommendation system. Meta page, Report Code.

• Improved mwtokenizer to accommodate non-whitespace languages and improve tokenization for languages with special sentence ending symbols.
• Improved performance of the link-recommendation system for 12 out of 23 failing languages by incorporating mwtokenizer.
• Addressed deployment bottlenecks and improved the Wikipedia link recommendation system by creating a language-agnostic model to replace the 300+ individual language-dependent models.
• Used Python, Sklearn, and Wikipedia2Vec to accomplish these tasks.

Data Analyst
Wikimedia Foundation

As a contract data analyst, I worked with the Search and Analytics team to analyze SPARQL queries along with Wikidata dump to help scale the wikidata query service. Analysis work is done using both Spark (Scala) and PySpark (Python) on Hadoop clusters. Analysis work: wikitech/User:AKhatun.

• Informed business decisions to reduce data import speed and query timeouts in Wikidata Query Service by analyzing Wikidata to find large and most frequently queried subgraphs. Analysis was done using PySpark and SQL.
• Extracted information from SPARQL queries and saved as a hive table to perform further analysis on it.
• Performed data analysis on Wikidata dumps and combined it with user queries analysis to identify the most frequently searched Wikidata subgraphs.
• Enabled regular Wikidata graph and query statistics monitoring by creating a pipeline using Airflow, Hadoop, and Spark that performs data analysis, calculates, and saves metrics regularly.

Outreachy Intern
Wikimedia Foundation

Selected as an Intern in Outreachy to work with the Abstract Wikipedia project under Wikimedia Foundation.

• Created a data pipeline to fetch Lua modules across all wikis and performed data analysis to identify important modules for centralizing in Abstract Wikipedia.
• Performed source code similarity analysis using unsupervised machine learning towards a more language-independent Wikipedia.

See more information about the tool, the methods of identifying important modules, and grouping similar modules in the MetaWiki page.

MetaWiki page: Abstract_Wikipedia/Data
Source Code: wikimedia/abstract-wikipedia-data-science
Phabricator Taks: T263678
Tool: abstract-wiki-ds.toolforge.org

Machine Learning Engineer
Therap BD Ltd.

Performed data analysis and applied machine learning algorithms on computer vision and time-series data for pattern recognition and prediction generation.

• Improved in-office attendance application with high-accuracy face detection in image and video footage.
• Analyzed production server logs to identify ideal downtime for application release. Used historical data to plot trends in usage and produced reports.
• Optimized care home efficiency by leveraging Google Vision OCR and hosted an AWS API to extract SpO2 and heart-rate measurements from images of pulse oximeters to enable swift detection of COVID-19 signs.
• Used sensor readings from sleep mats to implement a comprehensive sleep profile dashboard (sleep state diagrams, bed in/out statistics) of IDD individuals, thus helping caregivers tailor their assistance and receive emergency notifications.
• Increased care home safety and accessibility by developing an ML-based fall detection system using inertial sensor readings from smartwatches.

Research Assistant
SUST NLP Lab

Worked on developing larger datasets and implementing transfer learning based deep learning approaches for Authorship Attribution in Bengali Literature, thus far surpassing the existing systems. Work available in GitHub. Datasets available in Mendeley.

• Created multi-purpose language models with various tokenization methods and pre-training datasets.
• Developed deep learning architectures for authorship attribution in Bengali Literature. Finally overcoming the barriers of hand-drawn features, which were typical in this field, implemented transfer learning with language model pretraining to identify authors by analyzing their writing patterns. Also explored the effect of tokenization and pre-training dataset on the final task.
• Performed character-level text classification with mixture of recurrent and convolutional networks, additionally pretrained character embedding on several large datasets to improve performance.
• Curated and cleaned a large long-text authorship attribution dataset to assist in model development and evaluation without data acquisition costs.

Related Publications:

• A Continuous Space Neural Language Model for Bengali Language (ICCIT, 2019. IEEE)
• A Subword Level Language Model for Bangla Language (IJCCI, 2019. Springer)
• Authorship Attribution in Bangla literature using Character-level CNN (ICCIT, 2019. IEEE)
• A Transfer Learning approach on Authorship Attribution in Bangla Literature (ACM TALLIP, 2022)

Cheriton School of Computer Science, University of Waterloo

Master of Mathematics in Computer Science and Engineering

CGPA: 3.96/4.00

Advisor: Daniel G. Brown

Courses taken:
CS848 F22: The Art and Science of Empirical Computer Science
CS848 F22: Knowledge Graphs
CS889 W23: InfoVis for AI Explainability
CS889 S23: Value-Driven Technology

Shahjalal University of Science and Technology

B.Sc. in Computer Science and Engineering

CGPA: 3.89/4.00 (2nd in Class)

Completed undergraduate thesis on Authorship Attribution in Bangla Literature. Applied deep learning NLP techniques to achieve high performing scalable systems.
Advisor: Md Saiful Islam, Ayesha Tasnim
Thesis Report

Core Courses: Algorithm Design and Analysis, Data Structure, Database System, Object Oriented Programming, Software Engineering and Design Patterns, Technical Writing and Presentation, Artificial Intelligence, Introduction to Data Science, Machine Learning

Barbara Hayes-Roth Award for Women in Math and Computer Science

Received the Barbara Hayes-Roth Award for Women in Math and Computer Science for demonstrated academic excellence as a graduate student in University of Waterloo.

Outreachy Internship

Selected as one of 54 Outreachy Interns among 1000+ applicants through contributions in various Open Source projects.

Secure and Private AI scholarship

Secure and Private AI scholarship from Facebook. One of 300 out of 6000 candidates selected worldwide.

Best Research Poster Award

2nd Place, Best Research Poster Award, ICBSLP (International Conference). Presented our work on Authorship Attribution in Bengali Literature using transfer learning and compared it to existing systems and character-level CNN architectures.

Education Board Scholarship

Education Board Scholarship during undergraduate (4 years long) for best performance nation-wide awarded by the Education Board, Government of Bangladesh.

Programming Contests

• 5th in National Girls Programming Contest 2017
• 7th in National Girls Programming Contest 2018
• 8th in NSU Inter-University Girls Programming Contest, 2018
• ACM ICPC ASIA Regional Programming Contest, Dhaka Site 2018, 34th rank.

Directed Reading Program Mentor

Mentored a group at the Directed Readinig Program (DRP) where undergraduate students get introduced to new topics in Math and Computer Science and possibly some gentle introduction to research. My group learned about LLMs, ways to set up and use a personal LLM, and its applications in creative endeavors like story-telling.

SPARCS Workshop Panelist

Discussed my research and various opportunities in Computer Science with underrepresented students in Grade 9-10 across Canada. The aim was to bust the myths of Computer Science and invite inclusivity in the field, to show students how Computer Science is welcoming to all, whether you are math-savvy or not, tech-savvy or not, and what the recent career prospects look like for Computer Science students. Slides.

Student Volunteer at WiCS Technovation

Helped set up and guide young girls in high school and below during several Women in CS Technovation events at University of Waterloo.

Workshop Instructor at WiCS Con

Conducted a hands-on beginners AI workshop at University of Waterloo for the WiCS (Women in CS) Conference 2023. Attendees included Undergraduate and High School students. Attendees learned about AI and were given a run down on a simple ML problem using the Titanic Kaggle Competition.

Google Summer of Code and Outreachy Co-ordinator for Wikimedia Foundation

Organized events and helped applicants, mentors, and interns in all steps pertaining to the internships.

WikidataCon Panelist

Panelist for 2 sessions and a project presenter at Wikidata Conference (WikidataCon). Sessions, Video.

Abstract Wikipedia

Since my work during Outreachy Internship with Abstract Wikipedia, I have been working on improving and developing the abstract-wiki-ds tool to better perform clustering on source code. UI developement is also underway.
Taking small steps, but steps nonetheless.
Phabricator: T263678

Machine Learning Workshops

Conducted a series of IEEE beginners Machine Learning Workshop. Workshop materials available in GitHub.

Competitive performing trainer

Trained junior year undergraduate students for Competitive Programming.

Computer Vision Nanodegree

This nanodegree was the 2^nd phase of the Secure and Private AI Facebook Udacity Scholarship. Learned and applied Image Processing, Transfer learning, Kalman filters, Graph SLAM algorithm. Completed projects include Day-Night Detection, Facial keypoint detection, Object Detection, Image Captioning, Sentiment analysis and, Object Localization and Mapping.

Certificate

Data Scientist with Python

A bundle of Datacamp courses for python, data cleaning, manipulation and analysis, pandas, data visualization, SQL, statistical thinking and, machine learning.

Certificate

Practical Deep Learning for Coders

Jeremy Howard

Machine Learning & Deep Learning

Machine learning course and Deep learning specialization by Andrew Ng, Coursera. Thsese courses cover everything from the basics of machine learning and neural networks from scratch to deep learning techniques in computer vision and NLP with CNN, RNN, GRU and LSTMs, hyperparameter tuning and structuring machine learning projects.

Machine Learning Certificate
Deep Learning Specialization Certificate

​Oxford University 1millionWomentoTech ​Summer of Code

DIY (Do It Yourself) Track.
This summer of code was my introduction to Python and Machine Learning for the first tme ever. Tons of amazing volunteer mentors helped me a lot, from setting up python to understanding support vector machines, random forests. Setting up python and anaconda, in windows(!), was quite messy. Multiple python and too many incompatibality issues. I have a come a long way since then. Did a lot of assignments and solved ML problems with the completely hands workshops in this SoC.