Sep 2023 – Feb 2025 · 1 yr 5 mos · Mountain View, California, United States · On-site

• ● Designing XGBoost based ML model to detect scam tokens using last two years historical transaction details with help of AWS SageMaker to build, train & deploy model with 0.93 F1-score.
• ● Created an ETL pipeline using SQS and Redis to monitor price fluctuations in cryptocurrency tokens based on Ethereum Virtual Machine (EVM), utilizing this information as a signal to identify possible fraudulent activities.
• ● Built APIs and data pipeline to predict scam token using ML model in real time with time latency of 110 ms.
• ● Integrated Neo4j for real-time analysis of incoming transactions for potential fraud detection.
• ● Implemented continuous integration & continuous deployment (CI/CD) pipeline for daily system deployment & upkeep.
• ● Configured & managed Geth nodes of Ethereum, Polygon, BNB, Base, Arbitrum, Avalanche, Optimism, & Opbnb chains.
• ● Developed backend REST APIs with FastAPI & Flask to support dashboard that monitors scam & fraud alerts for clients.
• ● Technology Stack: Python, Pandas, Sklearn, Redis, Docker, DynamoDB, MySQL, SQS, SNS, Neo4j, SageMaker, & EFS

May 2022 – Aug 2022 · 3 mos · United States

• ● Analyzed users behaviors by visualizing users time spent, and number of sessions against various activities on Newsela.
• ● Applied PCA, random forest based feature importance and correlation analysis to identify significant activities and factors to increase user return rate and time spent on the platform using pandas, matplotlib, and seaborn.
• ● Clustered teachers by executing K-Means algorithm on user return rate, number of active days in school year, time spent & number of sessions and analyzed clusters to understand impact of prints by teachers on Newsela using scikit.
• ● Preprocessed data for clustering by filtering missing data, feature selection, normalization, and removing outliers.
• ● Extracted data from Snowflake using SQL query & python sqlalchemy to create user engagement metrics dashboard.
• ● Technology Stack: Python, Jupyterhub, Numpy, Pandas, Scikit, Snowflake, Matplotlib, Sqlalchemy, and Seaborn

Aug 2021 – Sep 2023 · 2 yrs 1 mo · Los Angeles, California, United States · Remote

• Lab: Integrated Media Systems Center(IMSC)
• Mentor: Dr. Seon Ho Kim, Associate Director, IMSC Lab, USC
• ● Applied computer vision to count number of Homeless Encampments on Los Angeles(LA) streets in real time using cameras installed on garbage vehicles of Los Angeles Department of Water & Power. (https://github.com/utkarshUSC/object_detection)
• ● Developed YOLO5(CNN based) model and detected homeless encampments with precision 0.86 & F1-score 0.83 on LA streets.
• ● Participated in IEEE BigData Cup 2022 to detect the road damage and its type where our team achieved 6th rank globally and
• fabricated ensemble of country specific YOLOv5 models. (https://github.com/USC-InfoLab/rddc2020/tree/crddc2022).
• ● Designed a semantic segmentation model using UNet to detect lanes and different lane marks in Taiwan’s traffic images.
• ● Technology Stack: Python, Pytorch, Fast AI, OpenCV, AWS Lambda, and Google Colab

Jan 2021 – Jul 2021 · 6 mos · Bengaluru, Karnataka, India

• ● Automated daily process for whitelisting & setting credit limits for more than 1 million users by creating distributed ETL data pipeline using golang, kafka, airflow, SQL and redshift.
• ● Enhanced existing REST APIs by increasing unit test cases coverage more than 75% and integrating in Github webhooks.
• ● Collaborated in launching Driver Pay Later product by creating REST APIs & Grafana dashboard to track performance.
• ● Migrated existing micro services with the help of terraform from Docker to Kuberbernetes.
• ● Technology Stack: Golang, Airflow, Kafka, Redshift, Redis, Docker, Git, Grafana, Jenkins, AWS, & Kubernetes

Jun 2018 – Jan 2021 · 2 yrs 7 mos · Bengaluru Area, India

• ● Led dynamic discounting project by implementing reinforcement learning based multi-arm bandit algorithm using A/B experiment which considers price elasticity, demand elasticity and competitor’s price to decide discounts on flights in real time to reduce burn $20K/daily and increase net margin $18K/daily.
• ● Coordinated with finance team to make real time redshift dashboard for monitoring effect of dynamic discount using SQL.
• ● Formed user and sector clusters using K-mean clustering according to their price and demand sensitivity in historical data.
• ● Built Random Forest & Gradient Boosted Tree models to predict flights’ fare in next five days with 83% accuracy.
• ● Trained multiple flight fare prediction models sector and air carrier wise on preprocessed data of last two years.
• ● Utilized correlation analysis, Chi square, visualization, and Lasso regression for feature selection of flight fare prediction.
• ● Performed hypothesis testing using T-test on recent data to decide update of model for flight fare prediction.
• ● Designed REST APIs, data pipeline, and deployed models to predict flight fare in real time within 50 ms using caching.
• ● Developed data pipeline which handles 10M requests/second for logging event data for each step of flight bookings.
• ● Contributed on REST APIs for flights search engine to produce relevant flight list according to user query in real time which handles throughput of 500K req/sec, & extract flight details from 3rd party vendors with avg. response time 100 ms.
• ● Maintained Flink job on EMR clusters to aggregate & store flight data on hourly windows for future data science projects.
• ● Technology Stack: Python, Golang, Numpy, Pandas, Scikit, Matplotlib, Seaborn, Jupyterhub, Pyspark, Flask, Django, Kafka, Redshift, Redis, Aerospike,Cassandra, Git, Linux, Jenkins, Flink, AWS, EC2, EMR, EKS, and S3

Jan 2018 – Jun 2018 · 5 mos · Bengaluru Area, India

• ● Participated in creation of recommendation engine to generate hotel ranks relevant to users & their queries using ‘Learning To Rank’ algorithm with NDCG value of 0.47 which increased booking to search ratio by more than 2%.
• ● Trained ‘Learning To Rank’ algorithm on previous four years data using ‘Pyltr’ library which executes LambdaMART.
• ● Designed APIs for offline evaluation and comparison of various ranking algorithms on historical data.
• ● Investigated significant features for hotel booking using correlation analysis, visualization, and exploratory data analysis.
• ● Wrote SQL queries to monitor impact of various hotel ranking algorithms in real time on the redshift dashboard.
• ● Technology Stack: Python, Scikit, Numpy, Pandas, Redshift, Airflow, Matplotlib, Linux, and Seaborn

Jun 2017 – Jul 2017 · 1 mo · Mumbai Area, India

• Project: Stitching of Microscopic Images (Using OpenCV & Python)
• Guide: Prof. Santosh Noronha (IIT Bombay)
• Applied and Compared different methods SIFT, SURF, BRISK, ORB, NCC and Mutual Information for stitching of microscopic images.
• Used DCT and gradient for quality assessment of stitched images.
• Accomplished stitching of upto 50 images with accuracy of 95%.

Sep 2016 – Jun 2018 · 1 yr 9 mos

• ●Fabricated and simulated trading strategies-Alphas using price, volume, fundamental, and analyst datasets with transcending sharpe & turnover requirements in real world simulation platform Websim on USA/EUR/IND regions.
• ● Adopted correlation analysis and visualization against returns to discover features and trading strategies for alpha creation.
• ● Applied statistical modeling and regressive models like linear, polynomial and random forest to invent robust alphas.
• ● Technology Stack: Python, Scikit, Numpy, Pandas, Matplotlib, and Linux

Jun 2016 – Jul 2016 · 1 mo

Jun 2016 – Jul 2016 · 1 mo

• Project: Automatic Number Plate Recognition (using OpenCV, Python)
• Guide: Prof. Gaurav Pandey (IIT Kanpur)
• Detected number plates of static vehicles taken from different angles under different illumination and colour level.
• In first step, used Sobel filter, Otsu thresholding and Morphological closing to get the region with high density of edges
• In second step, used height -width ratio and area of the region to detect plate area from them.
• Got an accuracy of 83% even in different environmental conditions.