Arbaaz Qureshi

Diablo Lake, North Cascades, WA, USA

Resume
email
GitHub
Twitter
LinkedIn
Google Scholar

I'm an Applied Scientist at Amazon, where I work with the Shopping Convo Foundation team on LLM-powered shopping agents and their evaluation. I'm a part of the core science team behind "Buy For Me", which lets customers buy products from third-party sites without leaving the Amazon app. My work there centers on making these agents trustworthy: reading product pages to pick and confirm the attributes a customer asked for, and building step-level verifiers that check the agent's actions against user intent before it takes any irreversible steps, like placing an order or adding to cart. Before Amazon, I was an Applied Scientist at Lowe's, where I built an agentic, multimodal conversational recommender for product discovery and comparison, and for answering home improvement questions. Here, I also built a personalized, session-aware query auto-complete system that is deployed on Lowe's website. Broadly, I'm interested in deep reinforcement learning and post-training large language models to better navigate complex, real-world environments.

I was a graduate student at the Robert and Donna Manning College of Information & Computer Sciences, University of Massachusetts Amherst, where my coursework centered on machine learning, natural language processing, information retrieval, and distributed systems. Along the way I worked with Prof. Mohit Iyyer on Research Paper Tagger, fine-tuning BERT-based classifiers to predict the research track of an NLP paper from its title, abstract, and authors.

I interned at Google in the Fall of 2022. Here, I worked with the Pixel Watch Ambient Compute team on the low latency off-body detection feature of the Google Pixel Watch. I built various convolutional neural network based detection algorithms to realize the said feature. During the summer of 2022, I interned at Lowe's, where I worked with Dr. Surya Kallumadi on extreme multi-label classification for semantic product search and query auto-completion.

Before moving to the US for graduate studies, I was a Research Fellow at Microsoft Research, where I worked with Sonu Mehta, Dr. Ranjita Bhagwan, and Dr. Rahul Kumar of team Sankie. I worked on source code processing to automatically classify and describe code edits in natural language, and on machine learning techniques which help prevent bugs and misconfiguration in large services.

Before joining Microsoft Research, I was a summer research intern at the GREYC lab, University of Caen Normandy, France. I worked with Prof. Gaël Dias on analyzing the effects of gender information on the estimation of depression severity using multimodal and multitask deep learning techniques.

I earned a Bachelor's degree in Computer Science and Engineering in 2019, from the Indian Institute of Technology Patna. I defended my Bachelor's thesis under the supervision of Dr. Sriparna Saha and Dr. Mohammed Hasanuzzaman, at the AI-NLP-ML lab, where I explored various multimodal and multi-task deep learning techniques to concurrently estimate depression severity and emotion intensity.

Amazon

December 2024 - Present

Applied Scientist | Manager: Dhruv Agarwal

Building and evaluating LLM-powered browser agents that accurately pick and confirm user-specified product attributes on product pages, as part of the core science team behind "Buy For Me", which automates shopping on third-party (non-Amazon) e-commerce sites.

Developed step-level verifiers that validate agent state against user intent and gate irreversible steps (add-to-cart, order placement), detecting deviations and triggering recovery by providing corrective feedback, cutting incorrect purchases by 93%.

Lowe's

July 2023 - December 2024

Applied Scientist | Manager: Surya Kallumadi

Developed an agentic multimodal, context-aware conversational recommender system using LLMs, which facilitates product discovery and comparison, and answers home improvement queries. This work was demoed to and closely followed by the CTO of Lowe's, Seemantini Godbole.

Worked with different LLMs (different GPT models, Gemini, LLaMA etc), and with different paradigms of in-context learning (few shot learning, chain-of-thought reasoning, retrieval augmented generation for home improvement project question answering) to realize an efficient chatbot.

Engineered a personalized, session-aware retriever and re-ranker for the query auto-complete system using text embeddings (transformers) and ANN indexes (HNSW, ScaNN). This system is deployed on Lowe's website.

The developed framework shows improved recall and a reduced average number of characters typed for the ground truth to appear in the auto-complete suggestions, resulting in a 2.2% engagement increase and $37M annual incremental revenue.

IBM

February 2023 - June 2023

Graduate Student Researcher | Managers: Dr. Taesung Lee and Dr. Youngja Park

Devised a method to obtain text descriptions per neuron, of concepts that activate those neurons in a BERT model. Our method alleviates the requirement of Human in the loop, to generate the text descriptions.

Created a dataset of 72K Amazon reviews and their annotated concepts, using various open-source LLMs like Flan-T5 XL and Pythia 12B. Analyzed which annotated concepts trigger neurons in the BERT language model using this dataset.

Work published at AAAI 2024 (ReLM workshop).

Google

September 2022 - December 2022

Engineering Intern | Manager: Cac Nguyen

The goal of this internship was to develop machine learning models which can detect whether a Google Pixel Watch is on-wrist, or whether it is off-wrist (low latency off-body detection).

Created an end-to-end framework which loads and processes wear data from the cloud, trains any generic neural network on the wear data to detect whether the watch is on-wrist or off-wrist, and evaluates how the network performs in a real-world setting.

Performed extensive experimentation with different convolutional neural networks and different off-body detection algorithms, and achieved a performance which competes with the deployed heuristic algorithm.

Lowe's

June 2022 - August 2022

Graduate Applied Science Intern | Manager: Surya Kallumadi

Trained extreme multi-label classification (XMC) models for semantic product search and query auto-completion, using the PECOS XMC framework.

The trained XMC semantic product search and query auto-completion models show recall values competitive with the deployed search and type-ahead prediction systems respectively, and (most importantly!) improve the per-query serving latency by at least 20 times.

Both the developed XMC models have been moved to production, and are serving approximately 20 million customer transactions every week.

Microsoft Research

August 2019 - August 2021

Research Fellow | Managers: Sonu Mehta, Dr. Ranjita Bhagwan, and Dr. Rahul Kumar

Worked on ML for code. Built various AI-powered source code processing tools: automatic code-edit classifier, and automatic code-edit description generator; these are deep graph encoding neural networks built on Code2Seq, that encode the abstract syntax tree of input source code [Paper].

Worked on association-rule mining based file recommender system which helps prevent bugs and misconfiguration in large services, where the association rules are mined from GitHub commits.

GREYC lab, University of Caen Normandy, France

May 2019 - July 2019

Research Intern | Managers: Prof. Gaël Dias

Worked on different multi-task deep learning networks and algorithms (shared-private and adversarial shared-private networks) to estimate depression severity separately in men and women.

Established that information about gender improves estimation of depression severity, and that adversarially learning to predict the depression score separately for men and women improves the performance of depression severity estimation in both the genders.

Works published in IJCNN 2021 and the French Journal of Psychiatry 2019.

AI-NLP-ML lab, IIT Patna, India

August 2018 - May 2019

Research Assistant | Managers: Dr. Sriparna Saha and Dr. Mohammed Hasanuzzaman

Worked on the multimodal fusion of audio, video and text modalities to estimate depression severity. Developed a novel attention fusion mechanism to achieve the SOTA results.

Worked on multi-task deep learning to concurrently estimate depression severity and emotion intensity. Established that learning to estimate emotion intensity supplements predicting depression severity.

Beat the then SOTA on the DAIC-WOZ depression dataset. Works published in IEEE CIM 2020 and IEEE IS 2019, and hosted on arXiv

Quantela

May 2018 - Jul 2018

Machine Learning Intern | Manager: Mr. Sanjiv Kumar Jha

Explored various sequence-to-sequence models like LSTM and GRU networks, to predict the local weather of Jaipur, India, based on different input parameters.

The trained seq-to-seq model exceeded Quantela's deployment performance.

Built time and memory profilers for Python functions as a side task.

Research Paper Tagger (RPT)
code | pdf
Advisor: Prof. Mohit Iyyer, UMass Amherst

Objective: Automatically tagging the research track of an NLP research article, given the title, abstract and the authors.

Built a dataset of 1744 research paper-research track pairs from ACL 2021, 2020 and 2019. Fine-tuned a BERT-based classifier on the collected dataset, on various combinations of title, abstract and the authors.

Achieved a top-1 accuracy of 70% and a top-3 accuracy of 83% on the test split of the collected dataset.

Cycle Location and Anti Theft System (CLATS)
code | pdf
Advisor: Prof. Jimson Mathew, IIT Patna

Objective: Tracking and preventing the theft of bicycles inside gated regions.

Built a working prototype of a lightweight system that detects the current location of the user's bicycle and prevents its usage without the consent of the owner.

Used RFID tags and readers (placed at strategically identified locations) to uniquely identify bicycles. A huge incentive for using RFID was the throwaway cost of an RFID tag to the end user (less than $0.15).

This project was chosen to be patented by the faculty of Computer Science and Engineering department, IIT Patna.

Towards generating informative textual description for neurons in language models
Shrayani Mondal*, Rishabh Garodia*, Arbaaz Qureshi*, Taesung Lee, Youngja Park
Association for the Advancement of Artificial Intelligence, 2024 (AAAI, 2024), ReLM workshop
website

Assessing the Effectiveness of Syntactic Structure to Learn Code Edit Representations
Syed Arbaaz Qureshi, Sonu Mehta, Ranjita Bhagwan, Rahul Kumar
arXiv preprint, 2021
website | pdf

Gender-aware Estimation of Depression Severity Level in a Multimodal Setting
Arbaaz Qureshi, Gaël Dias, Sriparna Saha, Mohammed Hasanuzzaman
International Joint Conference on Neural Networks, 2021 (IJCNN, 2021)
website | pdf | video | code

Improving depression level estimation by concurrently learning emotion intensity
Arbaaz Qureshi, Gaël Dias, Sriparna Saha, Mohammed Hasanuzzaman
IEEE Computational Intelligence Magazine, 2020 (IEEE CIM, 2020)
website | code

Multitask representation learning for multimodal estimation of depression level
Arbaaz Qureshi, Sriparna Saha, Gaël Dias, Mohammed Hasanuzzaman
IEEE Intelligent Systems, 2019 (IEEE IS, 2019)
website | pdf | code

The Verbal and Non Verbal Signals of Depression — Combining Acoustics, Text and Visuals for Estimating Depression Level
Arbaaz Qureshi, Mohammed Hasanuzzaman, Sriparna Saha, Gaël Dias
arXiv preprint, 2019
website | pdf

Automatic Prediction of PHQ-8 Questionnaire Scores using Artificial Intelligence
Gaël Dias, Arbaaz Qureshi, Sriparna Saha, Mohammed Hasanuzzaman
French Journal of Psychiatry, 2019
website

Amazon ‘Buy For Me’ Is The Latest Entrant In The AI Shopping Agent Race
Forbes, April 2025
link

Amazon can now buy products from other websites for you
The Verge, April 2025
link