CV

Mitacs Accelerate Intern, Mental Health Applications

Upbeing.ai

Designed and implemented a dialogue system for the Upbeing mental health application, contributing to research in AI-driven mental health support.

• Developed dialogue system for mental health support
• Applied NLP methods in wellness-focused applications

Senior ML Engineer, Social Media Applications

Arbisoft

Worked on personalized recommendation and information extraction for social media applications.

• Built personalized recommendation system for a dining-focused social media app
• Enhanced entity extraction accuracy on noisy short-text inputs

NLP Engineer, Banking Automation

ISSM.ai

Developed NLP-driven solutions to automate multilingual banking processes and customer support.

• Built Urdu/Roman Urdu transliteration and semantic similarity models
• Led chatbot development for United Bank Limited (UBL) and Allied Bank

NLP Engineer, Document Processing

Eighty.ai

Worked on document classification and information extraction systems for the financial and trade domains.

• Developed commodity classification system using HS codes across 97 categories and 1,244 subcategories
• Built high-precision NER and embedding models for the financial domain

Research Associate, Digital Forestry

Robotics and Intelligent Computing (RICE) Lab, LUMS

Conducted applied AI research for forestry management and conservation initiatives.

• Built monocular vision-based deep learning system for tree diameter estimation
• Created Pakistan’s first annotated forestry dataset adopted by WWF and Forestry Pakistan

NLP Engineer, Document Processing

Ai XPRT (AuditXPRT)

Worked on compliance and risk detection systems for insurance and banking clients.

• Developed anti-money laundering risk detection system for AIA Insurance
• Enhanced ComplyXPRT compliance system for Barclays Bank with lifelong learning, model distillation, and multi-threaded deployment

Research Assistant, Hate Speech and Generative Models

Knowledge and Data Engineering (KADE) Lab, LUMS

Contributed to research on hate-speech detection and generative models in Roman Urdu.

• Created first open-source Roman Urdu hate-speech dataset
• Developed custom detection model with improved Macro F1 score
• Explored differentiable surrogate objectives for GAN-based paraphrase generation

Lead Organizer

People's Climate March

Lead organizer for the New York City branch of the People's Climate March, the largest climate march in history.

• Awarded 'Climate Hero' award by Greenpeace for my efforts organizing the march.
• Men of the year 2014 by Time magazine

PhD

Dalhousie University

Computer Science

• Model Editing
• Machine Unlearning
• Interpretability
• Representation Learning

MS

Lahore University of Management Sciences (LUMS)

Computer Science

• Thesis: Hate-Speech and Offensive Language Detection in Roman Urdu

Doctoral Research Scholarship

Dalhousie University

Undergraduate Merit Scholarship

COMSATS University

Resolving Lexical Bias in Model Editing

International Conference on Machine Learning (ICML)

Current adapter methods are critically vulnerable to strong lexical biases, leading to issues such as applying edits to irrelevant prompts with overlapping words. This paper presents a principled approach to learning a disentangled representation space that facilitates precise localization of edits

Hate-Speech and Offensive Language Detection in Roman Urdu

Empirical Methods in Natural Language Processing (EMNLP)

Despite its prevalence, Roman Urdu (RU) lacks language resources, annotated datasets, and language models for this task. In this study, we: (1) Present a lexicon of hateful words in RU, (2) Develop an annotated dataset called RUHSOLD consisting of 10, 012 tweets in RU with both coarse-grained and fine-grained labels of hate-speech and offensive language, (3) Explore the feasibility of transfer learning of five existing embedding models to RU, (4) Propose a novel deep learning architecture called CNN-gram for hatespeech and offensive language detection and compare its performance with seven current baseline approaches on RUHSOLD dataset