Expert AI Data Trainer

• Developed and refined 150+ complex mathematical prompts to challenge Large Language Models (LLMs) and trained AI with step-by-step solutions in Linear Algebra, Discrete Mathematics, and Number Theory. Evaluated AI responses for logical consistency and improved problem-solving accuracy. Enhanced model reasoning capabilities, leading to better AI-generated solutions for complex queries.
• Assessed and debugged 500+ AI-generated responses, identifying and correcting 1,000+ inaccuracies, biases, and inconsistencies to improve AI reasoning. Designed structured assessment rubrics and validation criteria to ensure high-quality model outputs. These refinements led to a 20% increase in response accuracy and logical coherence.
• Collaborated with ML engineers and data scientists to optimize prompt creation, response debugging, and annotation processes, enhancing AI problem-solving efficiency.
• Managed AI training workflows using Slack and Jira, tracking hundreds of task tickets while contributing to cutting-edge advancements in NLP and machine learning.

Pro Account Sales Associate

• Built and maintained strong relationships with 20+ professional contractors, delivering tailored B2B solutions that enhanced client loyalty and long-term partnerships.
• Consistently exceeded monthly sales targets by 15–20% through strategic upselling, cross-selling, and promoting value-added services, leading to a 25% boost in repeat business.
• Awarded the 2024 Pro Excellence Award for driving a 34% sales growth and increasing average ticket size by 38%, demonstrating excellence in client engagement and revenue generation.
• Analyzed purchasing trends to forecast demand, optimize inventory, and ensure 95% product availability for high-volume professional clients, improving supply chain efficiency.
• Partnered with 10+ vendors and cross-functional teams to resolve client issues, achieving a 98% order accuracy rate and enhancing overall client satisfaction.

Graduate Research Assistant

Department of Computer Science, Irving K. Barber Faculty of Science.

• Led innovative research on Code Representation Learning under the guidance of Dr. Fatemeh Hendijani Fard, advancing methodologies for applying Large Language Models (LLMs) in code understanding.
• Developed and trained LLM-based models on datasets with 100,000+ code snippets, improving AI's ability to analyze and generate code and enhancing code understanding and representation capabilities.
• Collaborated with five peers on projects, including code comment generation, bug propagation detection, clone detection, code generation, and code translation, advancing research in software engineering.
• Created the largest dataset for code clone detection on PyTorch, compiling over 3 million Python code snippets from GitHub for PyTorch repositories, enabling robust training for clone detection models.

Graduate Teaching Assistant

Department of Computer Science, Irving K. Barber Faculty of Science.

Supervisors: Dr. Abdallah Mohamed, Dr. Scott Fazackerley and Dr. Firas Moosvi

• Managed and coordinated the Computer Programming I (Java) course for 250 students, implementing interactive teaching methods.
• Constantly received positive feedback from students for fostering a dynamic learning environment, enhancing overall engagement and course satisfaction.
• Contributed as a TA for Introduction to Data Analytics and Machine Architecture, supporting 60 students through tutorials and one-on-one assistance.
• Developed unit testing code in Python for the LearnCoding platform, improving testing efficiency and the overall learning experience for students.
• Generated iClicker questions from lecture slides, increasing in-class participation and reinforcing key concepts during lectures.

Part-Time Lab Instructor

Department of Electrical and Computer Engineering.

• Instructed over 400 students across courses, including Database Management Systems (MySQL), Programming Language I (C), Digital Logic Design, and Computer Architecture.
• Assessed term projects and conducted exams, providing individual guidance during office hours to promote academic excellence.
• Fostered a supportive learning environment, improving student performance through hands-on lab sessions and feedback.

Research Intern

• Partnered with Product and Food teams to analyze customer demand and forecast trends for Pathao, a top ride-sharing and multi-service platform in Bangladesh.
• Optimized shortest routes for delivery and ride-sharing, reducing travel times by 15% through analysis of 10,000+ data points.
• Designed an in-house mapping solution to replace Google Maps, saving $50,000 annually and enhancing route efficiency for ride-sharing, parcel, and food delivery services.
• Produced detailed analytics reports, improving service quality for over 1 million monthly users.

Undergraduate Teaching Assistant

Department of Electrical and Computer Engineering

• Assisted faculty in Data Structures and Algorithms, Computer Architecture, and Digital Logic Design, supporting course delivery and student engagement.
• Prepared course materials and conducted exam invigilation, enhancing course administration efficiency.
• Provided personalized counselling during office hours, contributing to an effective learning environment.

University of British Columbia, Canada

• CGPA: 3.40
• Completed Coursework: Advanced Database, Cloud Database, Computer Vision, Parallel Computing, Deep and Reinforcement Learning.
• UBC Graduate Research Scholarships - In Recognition of Excellent Academic Performances.
• University Graduate Fellowship - In Recognition of Excellent Academic Performances.

North South University

• CGPA: 3.82 (With Summa Cum Laude Distinction Award).
• 4^th, Innovation Challenge-8 competition - Capstone Project Showcase.
• 75% Scholarship on Tuition Fees - In Recognition of Excellent Academic Performance.

Exploring Code Clones in Software Development: A Study of PyTorch on GitHub and Stack Overflow

Master's Thesis | University of British Columbia

• Analyzed code cloning in PyTorch on GitHub and Stack Overflow, examining 10,000+ repositories to identify prevalent clone types and their distribution across deep learning stages.
• Curated a dataset of 3+ million Python code snippets using REST APIs and BeautifulSoup, contributing to a comprehensive resource on code cloning.
• Developed an in-depth analysis of how code duplication affects software development, providing valuable insights for improving code quality in AI models.
• Published thesis available at UBC Library.

A Blind Assistant Android Application Based on Yolo v5

• Employed YOLO v5, achieving 85%+ accuracy in real-time object detection, outperforming YOLOv3 for small image detection.
• Trained on a custom dataset (subset of MS-COCO with 10 classes and 10,000 samples) for 10 epochs on clustered GPUs.
• Developed a mobile app that uses real-time object detection to provide blind users with spatial awareness, identifying objects 3-5 meters away and delivering text-to-speech descriptions through Android’s TTS system.
• Improved usability by allowing users to identify objects based on proximity, location (left/right), and quantity within a 3-5 second interval.
• Prototype limitations: Struggles with detecting objects too close or too far away and managing information overload in environments with multiple objects.

Student Advisor System

• Developed a machine learning-powered advisor system, automating academic guidance for 500+ students.
• Integrated AI-driven models for course recommendations and personalized academic planning.
• Implemented CGPA calculators and graduation path visualizations, automating 80% of advisory tasks.
• Built the platform using Django, Tornado, jQuery, and Keras for seamless user experience and real-time recommendations.
• Enhanced academic decision-making efficiency with machine learning models for tailored guidance.
• Repository available at GitHub.

Bengali Automatic Speech Recognition with RNN-Transducer

• Pioneered Bengali speech recognition using the RNN-Transducer model, achieving a state-of-the-art Character Error Rate (CER) of 0.120% on the OPENSLR Bengali dataset.
• Developed a pretrained Bengali language model, reducing CER from 0.162% to 0.152%, and integrated it into the RNN-T framework for improved contextual understanding.
• Leveraged the Openslr English dataset by training a pretrained encoder, achieving a CER improvement from 0.152% to 0.143% for mixed-language recognition.
• Combined pretrained encoder and decoder models, leading to a significant CER reduction from 0.143% to 0.120%, enabling accurate recognition of both Bengali and frequently used English words.
• Overcame computational constraints to introduce continuous language testing, outperforming prior models limited to broken language formats, and setting a benchmark for future research in this domain.

Bengali Image Captioning with Attention and Pretrained Decoder

• Developed a state-of-the-art Bengali image captioning model using a modified ResNet152 encoder, attention mechanism, and pretrained decoder, achieving a corpus BLEU score of 0.547, surpassing previous benchmarks.
• Analyzed over 50,000 images and captions from the OpenSLR dataset to train a culturally nuanced model for both Indian subcontinental and Western contexts.
• Mitigated geographical bias in image captioning through tailored dataset curation, improving contextual accuracy in multilingual and multicultural scenarios.
• Validated the model's robustness with extensive testing, outperforming existing methods by a significant margin in descriptive Bengali image captioning.

Automatic English Text Summarization

• Developed an automatic text summarizer using Bidirectional LSTM with Residual Connection in a sequence-to-sequence framework.
• Evaluated on DUC2003 and DUC2004 news article datasets, achieving a 10-15% improvement in headline generation accuracy compared to baseline models.
• Generated concise, informative summaries for news articles, contributing to advancements in automated summarization technology.

12-bit Single Cycle RISC-based Processor

• Engineered a functional 12-bit single-cycle RISC-based CPU, supporting 16 unique instructions with a custom-designed Instruction Set Architecture (ISA).
• Developed a C++ assembler that processed 1,000+ lines of assembly code into machine code, ensuring seamless simulation.
• Designed and implemented the CPU's Data-path and Control Logic, achieving 100% operational accuracy during testing.
• Utilized C++, Assembly Language, and Logisim to deliver a processor simulation, reducing computation time by 20% compared to baseline models.
• Repository available at GitHub.

Food Canvas: Online Food Ordering and Restaurant Reservation Platform

• Developed a web application, "Food Canvas," enabling real-time restaurant reservations and processing 100+ simultaneous online food orders during stress tests.
• Designed and implemented a MySQL database supporting 50+ restaurant profiles and managing 10,000+ data entries for orders, reservations, and user interactions.
• Integrated front-end and back-end using the CodeIgniter framework, achieving a 25% improvement in query execution time through optimized database indexing.
• Enhanced user experience with responsive design using Bootstrap, JavaScript, HTML, and CSS, ensuring compatibility across 90%+ of devices tested.