May 2020 – Jun 2021 · 1 yr 1 mo

• Led a team of ML Engineers to build novel NLU capabilities for chatbots which power a Voice Bot architecture.
• Designed a pipeline to automate building a closed-domain bot given conversational audio recordings reducing bot building timeline from 6 weeks to 1 week using state of the art transformer models and clustering algorithms
• Built multi-intent handling capabilities by designing LSTM based multilingual sentence segmentation pipeline, achieves SOTA f1-score on mixed-code languages
• Integrated multilingual language models (XLM-RoBertA) into Rasa to build support for Indian languages understanding into chatbots without needing translation
• Developed framework to automatically test intent classification models for various biases

Oct 2018 – Apr 2020 · 1 yr 6 mos

• 1. Key creator and owner for Dekho App. I own the Data Science aspects of Dekho and am responsible for building the recommendation system powering it completely - From ideation, to experimentation, and data processing, to building a highly scalable ML system that can serve >2M users. The recommendation system uses a hybrid approach combining popularity and content based filtering.
• In Nov 2019, I have upgraded the system to use BERT (State of the art in Natural Language Processing) for content based filtering system.
• 2. Have worked on some smaller computer vision projects such as Depth map prediction to generate bokeh effect (using DeepLens paper approach), Content frame detection (detect which tv show / movie is being watched on streaming platform from the screenshot), TV Video Ad Classifier, and Face detection projects using both classical OpenCV approaches as well as deep learning based approaches (SSD, U-Net, Inception, etc).
• 3. Have given multiple presentations on the entire history of NLP till BERT, On Introduction to Machine learning, and on Recommendation Systems. Presentations attached herewith.

May 2018 – Sep 2018 · 4 mos

• Foyr is a real estate start-up which aims to put the process of getting a custom designed flat from the floor planning to interior designing on the web. I'm trying to lay down the foundation for the AI pipeline at Foyr. Currently building convolutional network (deep learning) solutions to detect and extract relevant entities like walls, doors, chairs, dimensions from a floor planning image, which would further integrate with their web app so as to offer an interactive experience to the user and designer.

Jun 2016 – Sep 2018 · 2 yrs 3 mos

• Authored an Artificial Intelligence based Logo Builder app that generates logos for businesses by brand name as input. Developed a font-pairing engine which utilized font vectors to output which two fonts would go together well. Genetic algorithms were used to improve the output and take user feedback in the loop. Flask was used to serve the Restful APIs and MongoDB powers the database of logos. Also designed a DSL to describe any design of a logo in a templated manner. Experimented a bit with GANs although it didn't work out well with them.
• Minimized the support ticket resolution time by developing and deploying a smart JIRA support ticket bot which finds similar tickets and assigns ticket to correct person. Doc2Vec and Tf-idf embeddings were used to train the model for similarity matching.
• Played a key role in building orchestration layer behind the WebPro panel through which web designers and resellers can maintain their businesses. Built RESTful APIs on top of springboot back-end. Applied Aspect Oriented Programming to develop an authentication layer, and added elasticsearch caching for instant search for orders at scale.
• Authored region based tax engine which can consume special taxes like GST, EU VAT. This made integration of any new taxes easy to the billing module of Orderbox - a control panel that powers businesses like BigRock, Hostgator, domain.com and over 5M customers.
• Tech Stack : Java, Python, Tensorflow, Gensim, Flask, Springboot, Elasticsearch, PostgreSQL.

Jan 2016 – May 2016 · 4 mos

• Upgraded Orderbox legacy code from Java 7 to Java 8 with 0 downtime. Orderbox is provisioning codebase behind thousands of web presence businesses, some of which are BigRock, Hostgator, Webhosting.info, domain.com, ResellerClub, etc.
• Integrated PayU Latam as a payment gateway for all 5M+ customers of Orderbox. Also secured other pre-integrated payment gateways by making them use TLS 1.2 for communication.

Sep 2015 – Dec 2015 · 3 mos

• Inductive Logic Programming is a relatively new field that lies at the intersection of
• logic programming and inductive machine learning. Inductive Logic Programming can
• construct rules or hypotheses by learning from examples and background knowledge
• something other fields of Machine learning do not do.
• My work involved finding out if I feed examples to an Inductive Logic Programming System, would it construct relevant features which could train a great machine learning model with maximum entropy. The study aimed to combine the logical power of Inductive Logic Programming,
• and Probabilistic wonder of Maximum Entropy model. It serves as a step
• towards the big goal of combining logical learning and probabilistic learning, which
• is together called Statistical Relation learning, or sometimes Probabilistic Inductive
• Logic Programming.

Aug 2015 – Dec 2015 · 4 mos

• Designed lab questions and managed the labs. Aided faculty in questions for projects and evaluation.

Aug 2015 – Dec 2015 · 4 mos · India

• Helped in making tutorial questions as well as lab questions and for handling the labs.

May 2013 – Jul 2013 · 2 mos · Greater Delhi Area

• Developed a responsive website based on Drupal CMS for 'Blood Bank' department of the organization, so as to help spread awareness about them, their camps, and their activities on the internet.