cv

Stanford University

Master of Science in Computer Science

Specialization: Artificial Intelligence

Distinction in Research

Cornell University

Master of Engineering in Computer Science

Cornell University

Bachelor of Science in Computer Science

Minor: Electrical and Computer Engineering

Summa Cum Laude (Highest Honors)

Aurora Innovation

Machine Learning Engineer (Perception)

Working on long-range perception for self-driving trucks

NVIDIA

Machine Learning Intern

Interned with the Perception team for autonomous vehicles

• Improved model training time for 3D detection models by 40%
• Devised methods to incorporate predictive uncertainty estimation for neural network classification and regression tasks

Uber ATG

Machine Learning Intern

Interned with the Perception team on 3D object detection

• Developed a new anchor-free 2D/3D image-centric object detection approach using camera + LiDAR sensor fusion
• Improved 3D object AP by ~3% and 3D object F1 Score by ~2% for pedestrian detection

Uber ATG

Software Engineering Intern

Interned with the Machine Teaching team on large-scale auto-labelling efforts

• Developed CNN to quicken 2D bounding box and segmentation labelling efforts using 3D LiDAR + camera labelled features
• Achieved ~15-20% improvement in labelling speed using model's pre-labels (based on preliminary experiments)

Stanford Artificial Intelligence Laboratory (SAIL)

Graduate Research Assistant

Worked with Prof. Stefano Ermon (in CS) and also with Prof. David B. Lobell and Prof. Marshall Burke as part of the SustainLab research group. Topics I researched: self-supervised learning, generative (diffusion) models, foundation models for remote sensing data, parameter-efficient training.

• Published multiple works on generative models, self-supervised learning and foundation models, including DiffusionSat, SatMAE, GeoLLM, Denoising Diffusion Bridge Models
• Awarded $100,000 Google Cloud grant by Stanford HAI for our research
• Mentored undergraduate students on research projects including data compression, parameter-efficient training, joint generative-discriminative models
• Gave a research talk at Stanford HAI Climate-Centered group

Cornell Computer and Information Science (CIS)

Undergraduante Research Assistant

Researched under Dean Kavita Bala and Prof. Bharath Hariharan on domain generalization in satellite imagery

• Devised an algorithm to tackle geographic generalization (i.e. robustness to variance in location of imagery)
• Implemented MAML, self-supervised learning algorithms and time-series vision models for multi-class crop classification

Cornell Data Science

Intelligent Systems Team Lead

Served as a team lead for Cornell Data Science, an undergraduate student-run project team. Was responsible for multi-semester projects, collaborations with industry, presentations to professors, and participation in competitions. Other team-management responsibilities included recruitment, onboarding, weekly team check-ins, planning social activities etc.

• Led a team to improve automatic mapping tools that can be used during humanitarian emergencies (project link)
• Implemented and validated object detection and semantic segmentation models to map houses in satellite imagery (precision: 85%, recall: 81%, baseline precision: 45%)
• Conducted interviews with 9 NGOs/organizations (Red Cross, Mapillary etc.) involved in humanitarian mapping
• Collaborated with IBM Research to access PAIRS satellite data and automate dataset generation
• Created a curriculum (assignments) spanning techniques in CV, NLP, RL and started a team blog to discuss research

CS 224W Class Final Project

Graduate Student

Final project for CS 224W: Machine Learning with Graphs

• Demonstrated how graph neural networks (GNN) can be used for online link prediction for drug discovery
• Featured as one of the best projects by Stanford CS 224W blog page (link)

CS 4300 Class Final Project

Undergraduate Student

Final project for CS 4300: Language and Information

• Developed a code search engine using embedding similarity search
• Runner up for best project

CS221: Artificial Intelligence- Principles and Techniques

Head Teaching Assistant

Led a team of 15 TAs to manage CS 221, a class with 250+ students (required for the AI-track)

• Led weekly meetings to organize TA duties, plan office hours and grading, and prepare homework assignments, exams, and section material

Stanford CS

Teaching Assistant

Served as a TA for CS 221 (Artificial Intelligence) and CS 161 (Algorithms) over the Spring 2022, Winter 2022, and Fall 2021 quarters

• Held office hours, led sections, prepared and graded exams & assignments, answered student questions on Ed

CS 4780: Introduction to Machine Learning

Co-Head Teaching Assistant

Served as one of the head teaching assistant for CS 4780 at Cornell over the Fall 2020 and Spring 2021 semesters

• Led sections, created demos, held office hours, prepared and graded exams, and answered student questions on Ed

Neurips 2023 Computational Sustainability Workshop

Reviewer

Reviewed paper submissions and recommended acceptance/rejection for presentation at the workshop

Neurips 2022

Group Leader- Education Outreach Volunteer

Raised awareness about AI to high-school students from New Orleans for Neurips 2022. Discussed ways to begin a career in CS and AI as well as some of the novel ways AI can change fields such as law, sustainability, science etc. While the students were curious to know what a day in the life of an AI researcher looks like, they were most excited about the free tech company swag :D

The Akanksha Foundation

Teaching Volunteer

Conducted educational activities and created worksheets to teach English and Maths to underprivileged kindergarten children

DiffusionSat: A Generative Foundation Model for Satellite Imagery

The Twelfth International Conference on Learning Representations

A novel diffusion-model based generative foundation model for satellite image datasets, that can condition on text and metadata. DiffusionSat can also solve inverse problems such as super-resolution, in-painting, and temporal prediction, surpassing the previous state-of-the-art

GeoLLM: Extracting Geospatial Knowledge from Large Language Models

The Twelfth International Conference on Learning Representations

A novel method to extract geospatial knowledge from LLMs using auxiliary map data OpenStreetMap. GeoLLM achieves a 70% improvement in performance relative to baselines

Denoising Diffusion Bridge Models

The Twelfth International Conference on Learning Representations

A more general framework to solve image-to-image translation tasks such as image editing. DDBMs outperform baseline methods by solving a stochastic differential equation based on the learned score of the diffusion bridge from data

Differentiable Weight Masks for Domain Transfer

ICCV 2023 Workshop on Out of Distribution Generalization in Computer Vision

Can gradient-based weight masking methods achieve domain transfer in computer vision models? We compare different weight masking methods to anaylze their effect on domain transfer in commonly used vision models

Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting

ICML 2023 Workshop on Knowledge and Logical Reasoning in the Era of Data-driven Learning

Chain-of-thought (CoT) prompting has previously been shown to elicit strong reasoning performance in LLMs. But does making the prompts logically invalid hurt LLM performance? Turns out- not so much!

SatMAE: Pre-training Transformers for Temporal and Multi-Spectral Satellite Imagery

Advances in Neural Information Processing Systems

A novel pre-training method for transformers for temporal and multi-spectral satellite imagery. SatMAE outperforms multiple prior state-of-the-art methods on a collection of datasets, demonstrating its value as a foundational model with strong downstream task performance

Neurips 2022 Scholar Award

Neural Information Processing Systems

Fully-funded attendance at NeurIPS 2022 conference, presented to a select group of student authors

Dean's List

Cornell University

Awarded in all semesters for strong academic performance

IBDP World Topper

International Baccalaureate

Achieved perfect 45/45 score (top 0.3% worldwide) in the IBDP grade 11-12 curriculum