I am a Computational Data Science grad student at Indiana University, Bloomington. I have around 6 years of software engineering experience, working on Search Relevance @ Flipkart (a Walmart company) and Data Science Platform at hyper-scaled startups. These teams process billion of events per day and are central to success of operational excellence of respective companies.
I am broadly interested in Applied Machine Learning and Data Science and building structural and semi-structural solutions to complex real world problems. In my professional life I tackled NLP and Time Series Forecasting.
My research at Indiana University focusses on Fairness Aware Graph Recommendation Systems where I am working on novel training methods for Graph Neural Networks.
In my free time I like spending time participating in ML competitions.
- Aug 2021 - May 2023Indiana University, Bloomington (Graduating: 1st Week May 2023)
MS (Computational Data Science)
GPA - 3.86/4.0
-
Jul 2011 - Jun 2015National Institute of Technology, Patna
B-Tech (Computer Science & Engineering)
CGPA - 8.32/10.0
- May 2023 - PresentIndiana University Network Science Institute
Working on novel model training methods to produce "Fairness Aware Graph Recommendation" models
- Aug 2023 - PresentKelly School of Business
Working as a paid RA on "User Intent as a Network". Project is a collaboration with Luddy and is funded by Kelly School of Business.
- Aug 2021 - May 2023NLP LAB @ IUB (Fall 2021)
Contributed extensively to design of TieML and Events' Timeline modelling
-
Jan 2019 - Jul 2021Swiggy
Software Dev Engineer II
Bengaluru, India
- Leading Data Acquisition Platform and Feature Store in Data Science Platform.
- DAQ is a tool used to capture, schedule APIs (used mostly for third-party APIs and dataset preparation). The architecture includes a proxy service, request generation module, transformation module, and scraping module.
- Feature store in DSP provides on demand features to DSP models in sub ms latencies on a very large scale.
- Worked on Forecasting and Correlation Platform which is used by teams to forecast concerned time series.
Sep 2017 - Jan 2019FlipkartSoftware Development Engineer
Bengaluru, India
- Involved in improvements in search intent model(CRF based) currently in production, identifying error classes, coming up with a solutions and fixing them.
- Worked on Fasttext based query store classifier, which predicts the category of a query, a new model using bi-lstms is going to replace this in near future.
- Wrote the first workflow to automate training and auto-deployment of various search models. First was written using Luigi and later migrated to Airflow.
- Was involved with various POCs with Data Scientists for search intent models.
- Wrote a generic framework using Airflow which at runtime creates generic dags for different ML models and orchestrates their training to deployment flow, including data and model validations.
- Used Cascading to extract data from various user events and then converting it to a form that can be used as training data for our models.
Sep 2016 - Sep 2017GrouponSoftware Development Engineer
Bengaluru, India
- Cyclops is the interface to Customer Reps which they use to resolve customer queries and requests which is live in all countries in which Groupon operates in.
- It talks to almost every microservice in Groupon and also exposes APIs to other internal services.
Aug 2015 - Sep 2016Netspeed SystemsSoftware Engineer
Bengaluru, India
- Being one of a handful of software developers participated actively in all development projects.
- Implemented many modules single-handedly (Virtual Channel Arbitration, Polarity based Arbitration, Multi-Cast Filtering, Structural Latency Breakdown, etc.).
-
Jan 2023 - Present
- Aug 2022 - Dec 2023
-
Jan 2022 - May 2022
Stocks Prediction (16th Rank AnalyticsVidhya)
This is a project based on competition held by AnalyticsVidhya.
Topic modelling (19th Rank AnalyticsVidhya)
In this contest solution, contestants had to come up with a solution to a multiclass text classification problem.
Hospitalization Period Prediction (115th Rank)
This was again a AnalyticsVidhya contest, where contestants were supposed to predict the period for which a patient is going to be hospitalized.
BiasNet: Learning to fight in StreetFighter II with induced Relational Bias from Differential Scenes
A novel deep on-policy model free actor critic reinforcement learning approach to act in a large action space using only the difference in scenes.
Bias Manifolds: Investigating Structure of Bias Manifolds and Bias Evolution
A survey of different approaches to study the structure of bias manifolds in different datasets. Also, a novel approach to study the evolution of bias in a dataset over a period of time.
BlindNet: Distilling world knowledge in Neural Networks
A survey of different possible neural network architectures to learn to understand the world using MS COCO dataset.
DeepFoodie: Clustering Food Items using Ingredient Embeddings
A novel approach to cluster food items using deep self supervised learning which uses ingredient embeddings.
Flipkart Hackday 9 Ekart Winner Hack 2018
Built a Multiclass model(inspired by Inception v3) to annotate images of lifestyle products. We also used Google OCR API to extract selected text from the tag. The end goal was to find top candidate FSNs. Text from tag was primarily used for features like price and brand. Others more important ones came from annotations(color, type, cloth type etc). On top of this to search that product we formed a query using this information and predicted using a CRF model trained on clickstream data of Flipkart using features from features generated using Flipkart’s catalog. It was so appreciated that it is in the process of going to production(which is the reason, not providing code pointer here). We did use differential learning rates to tune accuracies in the last stages of training to reach a 99.6% validation accuracy.
Groupon Geekon 2017
Deal recommendations to a user based on NSVD, using it as an unsupervised, collaborative filtering algorithm. Language: Python. Packages: Tensorflow, py2neo, pandas, and numpy. DB used was Neo4j. Dataset used was movielens dataset.