Selected Projects

For a complete list, kindly see my CV

HPV Chat Assistant demo

USC HPV Chat Assistant

Built an end-to-end RAG pipeline and chat assistant to promote HPV vaccination to USC patients using ChromaDB, FastAPI, OpenAI (text model and TTS model). Github Link

VeriSign signature verification demo

VeriSign

Developed a Siamese network of a combination of Convolutional, Max Pooling, Batch Norm and Dropout layers, to verify whether the signatures are signed by the same person or not using one-shot learning. Github Link

Chest X-Ray pneumonia detection

Chest X-Rays Pneumonia Detection

CNN was trained for detecting Pneumonia given Chest X-Rays. After training the model, test accuracy of 94.56% and a recall score of 0.97 was achieved.

Silatra sign language translation demo

Silatra (Sign Language Translation)

An Android application designed to translate the signs and gestures performed by the hearing- and speech-impaired into voice using Image Processing, Segmentation and Machine Learning algorithms KNN and HMMs. Paper published in IEEE.

Virtual Labs demo

Machine Learning Lab - IITB Virtual Labs

Contributed to the development of Machine Learning Lab on behalf of K J Somaiya College of Engineering, Mumbai and hosted on IIT Bombay's Virtual Labs website. As part of this, I developed simulations of neural networks, optical character recognition using Tesseract JS, Hebbian, Perceptron Learning Rules.

Experience

BILL

Aug 2024 - Present

Senior Machine Learning Engineer, Product AI

Spearheaded the company's first entity search system using document embeddings by working on pointwise LTR (Learning to Rank) algorithm and XGB ranker that achieves a top-1 accuracy of 91% using invoices and customer features. Retrieval is done against 27M embeddings and the RTT is 280ms on the OpenSearch VectorDB backend

Led the design of the intelligence layer of the LLM agents for various use cases including customer support, invoice field extraction while mentoring junior engineers and guiding other non-technical stakeholders on the use of the toolkit such as n8n, Braintrust, ElevenLabs, RetellAI and various prompt engineering techniques, evaluations

Contributed to a POC of multi-task learning on documents for field extraction using LayoutLM (2D attention transformer) embedding layer and Siamese network-based classification head achieving 89% classification accuracy and NER heads achieving 76% document field extraction accuracy

Contributed to the development of an end-to-end pipeline for field extraction for invoices using a rules-based engine to get low-latency and high-quality results. This pipeline achieves a p90 latency of under 350ms and a precision of 95%

BILL

February 2023 - Aug 2024

Machine Learning Engineer II, Network and Payments ML

Designed a proprietary logo extraction model from invoices using YOLOv8 model with a mAP of 0.9 and a similarity model - ConvNext with 87% accuracy to assist in identifying similar invoices with logos

Enhanced a document similarity model to improve de-duplication of merchants in the database using Siamese-BERT model which has increased the eligible transaction payment volume by 10% through enhanced payment rails

Information Sciences Institute, USC

January 2022 - December 2022

Research Assistant, May Team

Worked on designing bots capable of understanding communication and strategizing for the board game – diplomacy under the supervision of Dr. Jon May

Increased the win rates by an average of 5% for the powers in the game by generating and understanding DAIDE messages and using rules on top of the pre-trained reinforcement learning based Dipnet bot

Bill.com

May 2022 - August 2022

Machine Learning Engineer Intern

Generated a dataset of 1 million invoices and statements from user data after sanity checks for improving the amount OCR model

Leveraged an ensemble of document classification models to fix the imbalanced data distribution of statements in training dataset

Improved the amount prediction test accuracy from 75.5 to 77.6% with 50% confidence threshold

Information Sciences Institute, USC

February 2021 - January 2022

Research Assistant, Centre of Knowledge Graphs

Constructed a framework for identifying low quality statements in Wikidata knowledge graph amongst 1.1 billion statements on the basis of deleted, deprecated statements and constraint violations

Enhanced the graph embeddings of nodes using retrofitting based on BERT embeddings and structural, textual properties extracted from Wikidata, Probase and DBPedia datasets increasing Spearman correlation from 0.66 to 0.73 on WordSim353 benchmark

Barclays Global Service Centre

July 2018 - December 2020

Software Developer, Barclaycard UK

Devised a prototype fraud detection pipeline using Kafka queues, Cassandra DB and PySpark servers having an ensemble of ML models

Designed a real-time tweets sentiment analysis engine to enable quick customer service response achieving an accuracy of around 90 % in pilot runs

Created a classifier application utilizing ML algorithm LDA to extract insights from iOS and Android application reviews and customer complaints

Deployed a system that helps in connecting the colleagues with available bandwidth and skillsets with the colleagues needing assistance in their work, using AngularJS, Java, MySQL, saving more than 900 man-hours annually

Implemented dashboards for automated generation of real-time delivery metrics of more than 30 teams from Agile Central and Jira data sources which have been saving around 150 man-hours annually. Bagged the Barclays Award of Stewardship for this initiative

Virtual Labs, Indian Institute of Technology, Bombay

March 2017 – August 2017

Web Development Intern, Team Leader

Led a team of three to develop a Virtual Lab for the online demonstration of machine learning concepts such as neural networks, learning rules and optical character recognition

This lab has won the Global Online Laboratory Consortium International Lab Award

Publications

Concept similarity in Wikidata

We design a framework to determine concept similarity in large knowledge graphs such as Wikidata and Probase and experiment using language embeddings, graph embeddings, knowledge injections to determine the impact in enhancing the similarity scores.
Semantic Web Journal (October 2023)
PDF
Viola dialogue system

We present an open-domain dialogue system for spoken conversation that uses a topic-agnostic dialogue manager based on a simple generate-and-rank approach. Leveraging recent advances of generative dialogue systems powered by large language models, Viola fetches a batch of response candidates from various neural dialogue models trained with different datasets and knowledge-grounding inputs.
4th Proceedings of Alexa Prize (Alexa Prize 2020)
YouTube Team Info PDF
A Study of the Quality of Wikidata

We constructed a framework for identifying low quality statements in Wikidata knowledge graph amongst 1.1 billion statements on the basis of deleted, deprecated statements and constraint violations.
Journal of Web Semantics. Elsevier. (July 2021)
Code PDF
Creating and Querying Personalized Versions of Wikidata

We discuss several use cases where the KGTK kypher, a query language and processor can be used to execute various types of analyses on the full Wikidata KG on a laptop.
2021 Wikidata Workshop at ISWC
Code PDF
Does Wikidata Support Analogical Reasoning

We we investigate whether the knowledge in Wikidata supports analogical reasoning by using different types of text embeddings across different categories of relations defined in Wikidata.
Iberoamerican Conference on Knowledge Graphs and Semantic Web (2021)
PDF
Real-time Indian Sign Language Recognition

We designed an Android application which uses image processing, segmentation, HMM and KNN models hosted on a server to translate the gestures and signs of Indian Sign Language used by a hearing- or speech-impaired person.
9th International Conference on Computing, Communication and Networking Technologies (ICCCNT)
Code PDF
Keystroke dynamics analysis

I have explored the dataset for keystroke dynamics which could help in identifying patterns hidden in user's keystrokes. This could potentially help in performing non-intrusive real-time authentication of users. The analysis and models training has been covered over 2 parts.
Towards Data Science
Part 1 Link Part 2 Link Kaggle Notebook Link
LSTM back-propagation

I have detailed the derivation of the scary back propagation formulae of LSTMs in this Medium article.
Medium
Link

Certifications

Coursera logo

Natural Language Processing Specialization

Courses Included: 1) Natural Language Processing with Classification and Vector Spaces 2) Natural Language Processing with Probabilistic Models 3) Natural Language Processing with Sequence Models

DeepLearning.AI logo

Deep Learning Specialization

Courses Included: 1) Neural Networks and Deep Learning, 2) Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization 3) Structuring Machine Learning Projects, 4) Convolutional Neural Networks, 5) Sequence Models.

Kaggle logo

Kaggle Certifications

Links to All Courses: Intro to Machine Learning Intermediate Machine Learning Feature Engineering Pandas

Stanford Online logo

Machine Learning

Topics Covered: Logistic Regression, Artificial Neural Network, Machine Learning Algorithms, Principal Component Analysis, Collaborative Filtering

Contact

Please feel free to reach out to me for any query