Apr 2021 – Aug 2024 · 3 yrs 4 mos · San Francisco Bay Area

• Senior Machine Learning Engineer working in Siri Conversational Intelligence Team
• Orchestrated a team of skilled engineers to leverage contextual signals within Large Language Models to enhance Conversational Fluency and Disambiguation-related experiences in Siri and Apple Intelligence. Spearheaded highly collaborative efforts across cross-organizational teams. Role subsumed responsibilities below.
• Project Lead: Led a team to improve how Siri handles conversational follow-ups, using on-device coreference resolution to resolve ordinal, deictic and descriptive references [which comprises a part of the system published at CRAC 2023]. Responsible for the end-to-end project lifecycle, including architecture design, data collection, modeling, platform and internationalization. Successfully delivered user-facing features, fostering cross-organizational collaboration. Current owner of Siri’s entire ML-powered reference resolution stack.
• Modeling and Data Lead: Improved how Siri handles corrections and disfluencies using query rewriting.

Jan 2019 – Apr 2021 · 2 yrs 3 mos · San Francisco Bay Area

• Machine Learning Engineer working in Natural Language Modeling sub-team within Siri's Natural Language Understanding division
• Used on-device search and personal signals to move Siri’s parsing and contact resolution on-device in 13 locales, reducing errors by up to 20% (relative) over the server while simultaneously improving latency and privacy
• Applied state-of-the-art Named Entity Recognition and Normalization techniques to improve how Siri processes named entities such as movies and songs [published at NAACL 2021 Industry Track and Interspeech 2021], including via the novel use of speech/multimodal signals [published at 3rd NLP4ConvAI Workshop]

May 2018 – Aug 2018 · 3 mos · San Francisco Bay Area

• Machine Learning Intern with the Siri Natural Language Modeling Team
• Improved Named Entity Normalization by using Deep Generative Sequence-to-Sequence Models to learn commonly observed error patterns

Aug 2016 – Jul 2017 · 11 mos

• Added point-of-sale support for the FDI Global and Amex Direct gateways; Implemented the auto-auth reversal after void service for Amex Direct; Improved the CTP Sale API
• Designed and implemented a novel, near-zero technical debt run-time feature toggling framework
• National Winner, Visa Global Hackathon: Implemented an online infrastructure for seamless, one-click microtransactions using VISA Checkout APIs

Jan 2016 – Jun 2016 · 5 mos · Montreal, Canada Area

• Undergraduate thesis on Deep Learning in Computer Vision:
• Designed and implemented a novel Relative Face Pose Normalization and keypoint Depth Prediction with Recombinator Networks and a Spatial Transformer-like architecture without any ground-truth depths during either the training or testing stages, implemented using Theano and Lasagne
• Devised Convolutional Residual Memory Networks, a novel Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) Network based deep architecture inspired by Highway Networks and Residual Networks, both for Structured Face Keypoint Prediction with drastically reduced computational complexity (by a factor of 23x) and for Image Classification achieving near state-of-the-art accuracies
• Proposed Nested LSTMs, which outperform Stacked LSTMs on Character-level Sequence Prediction

May 2015 – Aug 2015 · 3 mos · Google Summer of Code

• Created a new Read-Evaluate-Print-Loop (REPL) Mode for Processing
• Worked on replacing the JEditTextArea with the more feature-packed RSyntaxTextArea, and on adding improved versions of several features (such as syntax and error highlighting and auto-complete support) into the new IDE

Dec 2014 – Jan 2015 · 1 mo · Bengaluru Area, India

• Worked on predictive customer scoring using Twitter feeds and other data signals such as information pertaining to firmographic data, company performance and so on, both with and without SAP HANA, and on integrating this predictive customer scoring system with Salesforce
• Proposed and coded up a basic predictive model, including natural language processing tasks such as tweet pre-processing, named entity recognition and sentiment analysis, as well as interfaces to Twitter, Salesforce and SAP HANA APIs using Java

May 2014 – Aug 2014 · 3 mos · Google Summer of Code

• Worked on the improvement of the Contributions Manager
• Added in new features such as a new type of contribution (examples-packages); enabled removal, install and update of contributions without restarting Processing; improved the UI of the Manager; improved how it handles the lack of an internet connection

May 2014 – Jul 2014 · 2 mos · Kalpakkam, Tamil Nadu

• Worked on the Application Oriented Benchmarking of a GPU Cluster using the Nanoscale Molecular Dynamics program (NAMD):
• Analyzed what combination of CPUs and GPUs provides optimum performance, and how the performance varies for different CPU-GPU combinations
• Analyzed how efficiently resources are used for a particular task and when adding an extra resource provides a speedup sufficient to warrant its addition
• Analyzed how different network architectures and protocols affect performance using the network monitoring tool Ganglia