cv

Ph.D. in Computer Science

McGill University, Montreal, Canada

B.Sc. in Software Engineering

Sharif University of Technology, Tehran, Iran

Visiting Researcher

ServiceNow Research, Montreal, Canada

• Worked in multimodal foundation model team and was supervised by a Senior Machine Learning Scientist
• Proposed a multimodal benchmark for evaluating state-of-the-art large multimodal models
• Designed efficient evaluation pipelines for large foundation models

Machine Learning Research Intern

Recursion, Montreal, Canada

• Worked in the Data Science group and was supervised by a Senior Machine Learning Scientist
• Designed models to learn gene representations for data-centric drug discovery
• Applied multi-modal models with self-supervised learning techniques to better integrate sequential and visual modalities of gene-perturbations to learn better gene representations

Research Assistant

University of Toronto, Toronto, Canada

• Worked in a group underthe supervision of Prof. Konstantinos N. Plataniotis
• Project goal was to improve the robustness of convolutional neural networks (CNNs) against adversarial attacks
• Implemented in Python 3

Software Engineering Intern

Moj Secure E-commerce Co., Tehran, Iran

• Interned in the AI team underthe supervision of the CTO and co-founder of Moj Company
• Project goal was to parse an arbitrary image of any valid bank card and detect the digits in the 16-digit client card number
• Used a Single Shot Detectorfor object detection and a CNN for digitrecognition
• Implemented in Python 3 using TensorFlow and Keras

PairBench Are Vision-Language Models Reliable at Comparing What They See?

• Developed a framework for evaluating Vision-Language Models (VLMs) as similarity kernels using four key metrics
• Demonstrated the importance of thorough assessment before adopting VLMs for evaluation tasks

WebMMU A Benchmark for Multimodal Multilingual Website Understanding and Code Generation

• Released WebMMU, a large-scale benchmark for evaluating multimodal and multilingual website understanding
• Benchmarked code generation and UI reasoning tasks across languages, highlighting limitations in current vision-language models

AlignVLM Bridging Vision and Language Latent Spaces for Multimodal Understanding

• Proposed AlignVLM, a framework to align latent spaces between vision and language models for improved multimodal reasoning
• Demonstrated that latent space alignment enhances cross-modal transfer and understanding across diverse tasks

Rendering-Aware Reinforcement Learning for Vector Graphics Generation

• Introduced a rendering-aware reinforcement learning method for generating high-quality vector graphics
• Showed improvements in fidelity and structural consistency compared to prior vector graphics generation techniques

BigDocs-7.5M A Large-Scale Dataset for Document Understanding

• Created an open-access dataset with 7.5 million multimodal documents for document understanding tasks
• Demonstrated that models trained on BigDocs-Bench improved document reasoning performance by up to 25.8% over GPT-4o

Guided Positive Sampling for Self-Supervised Learning

• Designed a novel method, namely GPS-SSL, for integrating a priori knowledge into any self-supervised learning (SSL) model
• GPS-SSL performs positive sampling by approximating strong augmentations
• The method encourages the base SSL method to be more robust against untuned data augmentations when applied to under-studied and/or real-world datasets

Revisiting Hotels-50K and Hotel-ID

• Revisited two image datasets, Hotels-50K and Hotel-ID, and proposed new training and evaluation splits with different levels of difficulty
• Proposed evaluation splits based on the images' class and super-class information to imitate real-world scenarios
• Accepted paper for the ICML 2022 DataPerf workshop

Structure-Aware Negative Sampling in Knowledge Graphs

• Design and implementation of a novel efficient negative sampling method with low computational cost for knowledge graphs
• Method based on considering the local neighborhood of each node when selecting the negative samples
• Accepted paper for the EMNLP 2020 conference
• Implemented in Python 3 using PyTorch

Scientific Paper Acceptance Prediction

• Design and implementation of neural network model which predicts acceptance of scientific papers with 84% accuracy, based on their abstracts and introductions
• Implemented in Python 3 using Keras and TensorFlow

Salary Trend Analysis

• Gathered and aggregated employee data from different occupation sectors from years 1996 to 2018 from Ontario, Canada
• Studied salary trends and found similar trend patterns exclusive to a few occupation sectors
• Used the Jensen–Shannon divergence between salary trends to construct a salary-similarity-network across employees in different sectors
• Built a multipartite graph from the employers and found meaningful employer embeddings

• Received Graduate Research Enhancement and Travel (GREAT) Award

• Received the Fonds de recherche du Québec – Nature et technologies (FRQNT) doctoral scholarship

• Ranked 3rd in DataJam Against Exploitation Canada competition

• Ranked 100th out of over 220,000 students in the National University Entrance Exam