I am an MSCS student at University of Southern California. Previously I worked in the Tech Division at Goldman Sachs in India and I interned at Google and Tessel@Rails Girls Summer of Code program. I am also an AI researcher, specialising in neural architecture search which is like AutoML but for convolutional neural networks. You can check out my publications here. At Goldman, I work in the SecDB Architecture team where we are working on internal trading databases in production. I have ample experience in working with databases, microservice architectures, APIs, production software deployment and maintenance and distributed systems. I completed my bachelors in CS from IIIT Delhi in 2020. I am also a core member of Women in Machine Learning and Data Science (WiMLDS) chapter in Delhi and have a bit of mentoring experience on my hands. I like to paint and watch Netflix during my free time.
This framework predicts the testing accuracy achieved by a convolutional neural network in one or more epochs. It helps in finding optimal neural networks, and reduces time involved in training candidate networks. Published under the guidance of Prof. Mayank Vatsa and Prof. Richa Singh (IIT Jodhpur, India)
Electrostatic precipitators (ESP) are widely used to capture fly ash generated due to combustion of coal in thermal power plants. This paper explains use of artificial intelligence to replicate human knowledge to optimize ESP performance and identifying the probable causes for many ESP related problems. Published under the guidance of Prof. Vikas Rastogi (Delhi Technological University, India)
Thesis project under professors Prof. Mayank Vatsa and Prof. Richa Singh to develop an algorithm that can predict neural network architecture for a given dataset, eliminating the need for hand crafted neural networks.
Self-explanatory as the caption is, in this project we sought to identify a dish, generate its recipe instructions and a preparation video from single image of a dish. We also created a dataset of recipes and their videos that we scraped. First, we identified recipe title of the dish in image, then used title and image to generate recipe instructions. Using recipe instructions, we generated an image corresponding to each instruction using DCGAN and DALL-E networks. Putting these images together gave us the preparation video.
Content retrieval models, multimedia cross-modal models, neural networks, attention networks, trans-formers, image-to-text, text-to-video, generative adversarial networks (GANs)We know regex isn't exactly easy to get correct, right? What if a DL model could give you the regex when you input a simple description? We used nearest neighbors and semantic unification as baseline models, then we proposed the Conv Seq2Seq model for this task. For testing, we use DFA equivalence to determine if two regexes are same.
Natural Langugage Processing, Attention Networks, RNN, GRU, LSTM, Autoencoder networks, Nearest Neighbors, DFA, State Machines, Semantic UnificationExploratory independent project to understand different video question answering models and datasets, and devising improvements and new models for the task
VideoQA models, Nearest Neighbor, Two-Stream Model, Faster RCNN, Word2VecNAS is automating designing of convolutional neural networks. This was previously a long trial-and-error driven task, involving long iterations of testing different candidate networks. The search space of possible neural networks for given depth is also large. Different methods like reinforcement learning, genetic learning, hyperparameter optimisation have been used for NAS but these incur huge computational costs. My work primarily focuses on making NAS computation effective so that it can be run on single GPU in a realistic time frame.
Reinforcement Learning, Genetic Learning, Hyperparameter Optimisation, Deep Learning, Neural Networks, Performance Prediction, Database CharacteristicsCancer cells can be detected from images by identifying cell nuclei in images. Image segmentation models coupled with post-processing techniques give a good accuracy (measured by IoU metric). We used Mask RCNN model backed by ResNet101 for image segmentation.
Image Segmentation, Mask RCNN, ResNet, IoU metric, Post Processing, Dataset Augmentation, KaggleIt has been observed that each sleep stage has some distinguishing features irrespective of the person. We can use this to identify a person's sleep into five stages. We extracted features like kurtosis, variance, mean of voltages that helped us train models better, and yield better classification accuracy.
Feature extraction, PCA, signal based features, SVM, ANN bagging, ablative analysisWe got a dataset of 150 features and 9557 raw samples from Kaggle. After cleaning the data and augmenting the dataset with SMOTE augmentation, we compared different baseline models and proposed a voting classifier model consisting of four classifiers.
SMOTE Augmentation, Voting Classifiers, SVM, Dataset Preprocessing, Ablative analysis, AdaBoost Classifier, Logistic RegressionIn this project, we develop a deep Q-learning RL agent for playing Atari games. The agent learns a policy to decide which action to play, given a state, so that the cumulative reward is maximized. We applied exploration vs exploitation policies wihth DQN network to train RL agents.
Reinforcement Learning, reward function, Atari games, learning rate, parameter tuning, pacman, breakout, exploration vs exploitationDeveloped on a driving assistant system that can identify lanes, road signs and objects on road. We used sobel filters, hough transform and other techniques to identify lanes. For road signs, we preprocessed the image to highlight road signs and then fed frames to SVM classifier to identify correct road sign. We used YOLOv3 model to identify objects like cars, trucks etc. on road.
Computer vision, object identification, edge detection, YOLO model, image preprocessing, SVM classifierI worked on this project during my research internship at Delhi Technological University in 2017. We wrote a procedural AI to predict faults using ML models, like current or voltage fluctuations, corona buildup etc. and provide accurate diagnosis. This project was developed in conjunction with on-field experts of ESP systems.
Feature selection, prediction models, industrial equipmentWrote a program in C to run multi-client chatroom, where poeple on different devices could connect to a central server and send messages to all the other clients.
Socket Programming, Operating Systems, C LanguageBuilt a Java program simulating flight booking system with implementation of mutex locks on flight database while spawning multiple reader and writer threads.
Databases, Synchronisation, MutexBuilt a complete chain reaction app using Java with an AI one-player mode. The AI player would use heuristic algorithm to select its next move. We used JavaFX for graphics and used multithreading concepts intelligently to run the game algorithm.
JavaFX Graphics, Multi-threading, Algorithms, Design Patterns, Java Programming, AI heuristic strategiesBuilt an obstacle avoiding robot from scratch using Raspberry Pi, motors and ultrasonic sensors. The robot can navigate its way around an obstacle course on its own.
Raspberry Pi, microcontrollers, sensors, Python programming, calibration, roboticsBuilt a complete tic-tac-toe game in Python with an AI one-player mode. The AI player used strategies and history of player's previous moves to select its next move. I have actually lost a few games against it :p
Python graphics, AI heuristic strategies, algorithmsBuilt a system to automate meal coupon collections using student ID cards in hostel mess, complete with a website UI built on LAMP stack and a Raspberry Pi 3 with RFID reader.
Raspberry Pi, microcontrollers, RFID, sensors, LAMP stack, frontend, backend, website UIBuilt a Python app to efficiently run multiple plagiarism tools on student submissions and view their result to aid professors in their courses. The app allowed for code comparison in multiple langugages and served as a unified interface to run different tools and highlight plagiarism cases.
Algorithms, programming, plagiarism toolsPresented my NEAP-F paper in AAAI 2021!
NEAP-F selected to be published in AAAI 2021! This paper was based on my bachelor's thesis work on predicting performance of neural networks.
Formally graduated from IIIT Delhi, obtaining Bachelors in Technology with Honors.
Started working at Goldman Sachs in WFH
8th semester finally over!
• Started working on recipe generation project. Using a single image of a dish, we can identify the dish name, generate its preparation instructions and video too!
• Volunteered for college cultural fest Odyssey. Quite a learning experience on crowd management and cold calling sponsors
• Started working on regex generation project. The aim was to provide a natural language description as input and output corresponding regex. Got to read papers by Regina Barzilay!
• Started working on independent project on video question answering systems
• Joined Women in Machine Learning and Data Science as core member :)
• Got elected as student council member in college
• Got job offer from Goldman Sachs
• Completed 16 week internship stint at Google India with Cloud Search team
Started working on bachelor's thesis on neural architecture search!
Completed a hectic semester with artificial intelligence, machine learning and image analysis courses
Attended GoLab conference in Florence, Italy, thanks to Rails Girls Summer of Code!
Completed 12 weeks internship at Tessel with Rails Girls Summer of Code
Completed 80 hours of teaching basic maths and computer science to 9th standard students at Patratu School of Economics
Attended Facebook F8 conference in California, US. My first industry experience!
Presented my first research paper in NFEST 2018!
My first paper on AI fault diagnosis system for electrostatic precipitator selected to be published in NFEST 2018
Completed research internship at Delhi Techonological University, worked on fault diagnonsis systems for electrostatic precipitators
•Built my first AI tic-tac-toe bot!
•Built obstacle avoiding robot using Raspberry PI 3
Joined IIIT Delhi as an undergraduate in computer science