Hey, I'm Utkarsh!

A passionate Software Developer experienced in software engineering at scale, distributed systems and machine learning.

What I bring

Software Engineer interested in distributed systems and machine learning at scale.

C++
C#
Java
python
C
JavaScript
AWS
Azure
Apache Storm
git
PyTorch
scikit-learn
Jupyter
Windows
Linux
Unix
docker
NoSQL-DBs
RDBMS
SQL

3 years of industry experience from Microsoft in large scale distributed systems, information retrieval and ranking

Master's in Computer and Information Science from University of Pennsylvania, with a focus on distributed systems and machine learning

Knack of working on 0-1 products and open-ended problems through prior experience

Education

University of Pennsylvania

Master of Science in Computer and Information Science

Jan 2021 - May 2022

Research in Bayesian Machine Learning, Reinforcement learning and projects in Distributed Systems

Ongoing research in adversarial attacks on reinforcement learning used in large scale software systems
Ongoing research in optimization of high dimensional black box functions using Bayesian Machine Learning
Built an end to end search engine from scratch on 500k webpages
Built a fully functional, distributed clone of GMail and Google Drive from scratch
Teaching Assistant for Machine Learning in Fall 2021 and Spring 2022
Coursework in distributed systems, machine learning, deep learning
GPA: 4.0/4.0

Manipal Institute of Technology

Bachelor of Science in Computer Science

Aug 2014 - May 2018

Ranked top 20% in the program. Took courses about Software Engineering, Web Security, Distributed Systems, Machine Learning

Minor in intelligent systems with coursework in Machine Learning, Natural Language Processing and Social Network Analysis
Took systems courses like Operating Systems, Distributed Systems, Networking
Built an automatic code-summarizer for Java Code using CNNs with attention mechanism
Participated in research of detecting glaucoma from high resolution retina images using CNNs
GPA: 9.11/10

Experiences

Research Assistant

Sep 2021 – Present

Adversarial attacks on Reinforcement Learning Algorithms deployed in production Software systems

I am currently researching and building novel attacks on reinforcement learning algorithms like PPO and A2C in systems environments like load balancing, congestion control, adaptive streaming etc.
The goal here is to make these algorithms perform worse than a greedy strategy.
Simulateneously, I also investigate fail-safe algorithms for shielding against such attacks, to guarantee at-least greedy performance.

Teaching Assistant

Sep 2021 – Present

Teaching Assistant for the graduate Machine Learning course for Spring 2022 and Fall 2021

My responsibilities are designing and grading homeworks and midterms, as well as helping students understand the content better.

Research Assistant

Aug 2021 – July 2021

Software development for DCIST: A fully autonomous, heterogenous swarm of robots to assist humans in combat scenarios

Built an automated metric-collection and visualization module from scratch for each simulation. Worked with stakeholders in identifying key metrics to collect and visualize.
Developed a batch-simulation mechanism. This allowed other researchers to specify several variable values in a YAML file, and then automatically run simulations for different combinations of values, hence streamlining productivity.
Implemented automatic aggregation and filtering for batch-simulation metrics.
I also researched new heuristics for probabalistic path planning of robots.

Software Engineer 2 & 1

June 2018 – December 2020

Division: Microsoft Search, Assistance and Intelligence

From June 2019 - Dec 2020, I worked in Bing Enterprise Search, where I led the design and development of a new backend infrastructure for serving acronym search results. The new infrastructure finally saw 30% gains in latency and ~5 percent gain in availability. This was while serving about 500M - 1.5 billion requests across all endpoints.
From June 2018 - June 2019, I was with the Microsoft Assistant Notifications team, where I was an active part of building C# microservices and library packages to power a new microservices-based Notifications system serving 10s of millions of users per day.
I also routinely took part in Site Reliability Engineering, mitigating, investigating and root-causing failures encountered on production systems.

Software Engineering Intern

Jan 2018 – June 2018

Division: Microsoft Search, Assistance and Intelligence

I built a prototype of a microservices-based Notifications system for Microsoft Assistant (previously Cortana) from scratch. The earlier system was centralized.
The new prototype offered more scalability, modular deployments and easier debugging.
The work from my internship laid the ground work for moving the Notifications system to a microservices-architecture, that I was an active part of.

Software Engineering Intern

May 2017 – July 2017

Division: Microsoft Search, Assistance and Intelligence

I built a voice controlled skill for Cortana to add-remove items from Cortana Lists using voice command
I also built a JavaScript Chrome extension to automatically parse out the lists from any given webpage and add it to Cortana Lists

Personal Projects

Some personal projects that I have built for fun or coursework, often both.

PennSearch

A google like end-to-end search engine deployed on cloud built from scratch

Crawled over 500k+ webpages using a Mercator style crawler using Apache Storm and deployed on Amazon EC2
Built a forward and reverse search index using Amazon Elastic MapReduce on over 2 million entries
Used a weighted combination of TF-IDF and PageRank for ranking of search results
Used Vue.js for UI
Backend written in Java using SparkJava HTTP Webserver package

PennCloud

A fully distributed GMail and GDrive Clone built from scratch

Powered by a distributed GFS like key-value store using active log-based replication and log-based fault recovery
Drive Storage features include creation, uploads, downloads, rename, moving of files and (nested) folders
Mail serving supports email functionality including send, forward, reply, reply-all
In-built load balancing functionality on both application and storage servers
Technologies used: C++, JavaScript, HTML5, Material CSS

TSsBO

High dimensional Bayesian Optimization of black-box functions by learning the active subspace

The objective is to optimize high dimensional, black-box non-differentiable functions by learning the active-subspace of these functions.
We learn a low-dimensional manifold of a high-dimensional function where Bayesian Optimization techniques like qExpectedImprovement and Thompson sampling can find local and global optimas, which can then be used to locate the optima in the original dimensions.
Technologies: PyTorch, GPyTorch, BoTorch

Online Certifications

Online Courses that I have completed