I enjoy building bleeding edge consumer facing products that are made possible by new advances in machine learning. I consider myself both a scientist and an engineer, and do not shy away from either invention or implementation. On the research side, my interest lies mainly in the intersection of representation learning and low resource machine learning - on how to design algorithms that can better deal with noisy or sparse datasets, or more generally, on constrained problem settings that are more likely be encountered in real life scenarios - this has led me to work on projects in model compression and acceleration, zero-shot learning and weak supervision techniques.
On the engineering side, production code that I have written currently powers online ML model building for two different services spanning Alexa and Amazon Web Services - namely, the Alexa Skills Kit (ASK) and AWS's Amazon Lex spoken language SDKs. I have been lucky to have the chance to work on state of the ML research as well as software architecture, and have over half a dozen patent applications pending with the USPTO in fields such as deep learning, NLP, question answering, spoken language technology and recommender systems. Resume (last update: 19 October 2017)
In a Nutshell
Alexa Brain, ASK ML teams.
Advisors: Michael Collins and Peter Bellhumeur.
“Zero Shot Learning Across Heterogenous Overlapping Domains”
International Conference on Spoken Language Processing (INTERSPEECH) 2017
Anjishnu Kumar, Pavankumar Muddireddy, Markus Dreyer and Bjorn Hoffmeister
"Just ASK: Building an Architecture for Extensible Self-Service Spoken Language Understanding"
1st Workshop on Conversational AI at the Conference on Advances in Neural Information Processing Systems (NIPS) 2017
Anjishnu Kumar, Arpit Gupta, Julian Chan, Sam Tucker, Markus Dreyer, Bjorn Hoffmeister, Ankur Gandhe, Ariya Rastrow, Denis Filmonov, Stanislav Peshterliev, Christian Monson, Agnika Kumar
"Efficient Large-Scale Domain Classification with Personalized Attention"
Accepted for publication at 56th Annual Meeting of the Association for Computational Linguistics (ACL) 2018
Young-Bum Kim, Dongchan Kim, Anjishnu Kumar, Ruhi Sarikaya
Natural Skill Interaction
Designed, prototyped, and helped productionize system that enables customers to talk to Alexa skills without needing to remember the name of the skill. Watch Alexa Science VP Rohit Prasad describe the feature at AWS Re:invent 2017.
Alexa Skill Suggestions
Built the first version of a system used to recommend relevant Alexa Skills to customers.
Amazon Lex: Spoken Language Understanding as an AWS Service
The work done for ASK was developed and deployed as a spoken-language understading service for AWS, known as Amazon Lex. Lex can be thought of as ASK minus Alexa.
Alexa Skills Kit: Spoken Language Understanding Platform
Built the foundational capabilities for spoken language understanding platform that underpins the Alexa Skills Kit, which has deployed over 25,000 skills in the US, 30,000 all over the world.
Conditional Random Field Model Compression
As an intern on the Amazon Echo team, I used a combination of voronoi iteration, bloom filters, feature hashing, elastic net regularization and weight quantization to reduce model sizes by 95% and increase inference speed by 25%, without any statistically significant loss in accuracy.
A NEURAL LATENT VARIABLE MODEL
Patent approved by USPTO, Publication Date - Februrary 2018
Inventors: Anjishnu Kumar and Markus Dreyer
- Note: 8 more patent applications are currently pending with the USPTO: Watch this space.
ASK Alexa PyKit
A minimalist Python library to build Alexa Skills using AWS Lambda. This library used to teach CIS 700 Deep Learning for Automated Discourse at the University of Pennsylvania.
Ask Alexa : Twitter
An Unofficial Twitter Alexa Skill written using ask-alexa-pykit.
I implemented an Alexa Skill for Twitter, after pitching it to their business development team - it was adopted by a software engineering team as Twitter's official skill for Alexa.
SUpeR REsolution, (A) Library. Implemented using Deep Convolutional Generative Adversarial Networks (DCGANs) in Keras.
Contributions to MxNet's Gluon API, including advanced activation functions and examples.
A local social network Android App with Machine Learning powered event mining from Twitter.
A machine learning augmented image gallery for Android for semantic visual search. Basically Google Photo's image search implemented using a cloud-side cascade of Support Vector Machines, a year before Google launched the feature.
Crackr: Supervised Keyphrase Extraction
A supervised keyword extraction system using a combination of Brown Clustering and RAKE. Given a small training sample, can learn to extract similar keyphrases. Developed for CollegeFeed to extract valid phrases from job listings. E.g. recognize Rust, Scala as being important keywords given Python and C in training data.
Machine Learning Scientist, Amazon Alexa Knowledge
Cambridge UK, September 2017 onwards
Semantic Parsing, Knowledge Bases, Semantic Search.
Machine Learning Scientist, Amazon Alexa Machine Learning
Worked on the Alexa Skills Kit and Alexa Brain initiatives.
Master of Science in Computer Science, (Applied Machine Learning)
University of California, Berkeley
Electrical Engineering and Computer Science (Exchange Student)
Birla Institute of Technology and Science (BITS), Pilani
Bachelor of Engineering (With Honours), Chemical Engineering
Internships and other Exposure
Data Science Intern, Opinion and Sentiment Mining
Built a prototype opinion mining system using Recursive Neural Networks.
Machine Learning Scientist Intern, Amazon Echo
Cambridge, Mass., USA
Used machine learning and probabilistic data structures to compress a ML model by 95% and reduce inference time by 25% with no significant loss in accuracy.
The Earth Institute, Data Engineering Intern, Project SharedSolar, Sustainable Innovations Lab
I worked on data engineering tasks for a distributed solar array in northern Africa, working on anomaly detection in sensor readings, with an aim to develop dynamic pricing strategies in the future.
Electronic Arts, Software Engineering Intern, Digital Advertising Platform
As a server side engineering intern, I migrated EA's push notification serving architecture from a monolithic to a distributed software architecture. I also built the components that serve push notifications to Amazon's Kindle line of devices.
Weekly Columnist, wrote a column on Software and Technology for 3 years.
Collegefeed, Consultant, Head of India Business Development, Part-time NLP Software Engineer
Collegefeed was an early stage startup specializing in recruitments targeting college student. I donned several hats there, drafted growth strategies to penetrate the Indian market and executed customer acquisition campaigns in top tier Indian and US colleges. I also invented/implemented a natural language processing library that automatically discovers keywords from an employee's resume that are relevant to any given job listing.
Management Consulting Intern, India e-Governance Advisory ,
I worked as part of PwC's advisory division, helping draft policy and technology strategy for the Government of India's GST (Goods and Sales Tax) project, the largest tax reform in the world.
National Thermal Power Corporation of India, Process Engineering Intern ,
As a process engineering intern in the Coal Processing division, I worked on regression models to project possible monetary savings and efficiency gains in a coal waste re-use project.
Business Analyst Intern, Business Development / Strategy.
Created an assessment report projected the growth of wireless internet technologies in India through 2020, the report was used to pitch to the head of Apple India.,