San Francisco Bay Area
Passionate about the intersections between Computer Science and Mathematics. Currently exploring Machine Learning and loving it. You can reach out to me at [email protected]. Microsoft Research profile: https://www.microsoft.com/en-us/research/people/apurvagandhi/
Architect and researcher for Microsoft Copilot working on leveraging program synthesis techniques for LLM orchestration, natural language commanding and plugins.
I am part of the Office Machine Learning (OXO ML) team, helping infuse AI features in Microsoft Office applications and surfaces. I also collaborate with the Gray Systems Lab (GSL) on the Tensor Query Processor: https://arxiv.org/abs/2209.04579. I created the Office Domain Specific Language (ODSL) (https://arxiv.org/pdf/2306.03460.pdf) and the original prototype for Office Chat Copilot (now M365 Copilot). I lead ML architecture for natural language commanding for M365 Copilots. With the help of my teammates, leadership and cross-org partnerships, ODSL has grown to support natural language commanding for all core office apps, many M365 apps, PowerApps and more.
Microsoft AI Development Acceleration Program (MAIDAP)
Worked on the Amazon Lex team. Lex builds conversational interfaces for your applications powered by the same deep learning technologies as Alexa. 1. Researched and prototyped improvements to Lex's NLP using Neural Text Normalization. 2. Designed and implemented three new customer-facing APIs to allow users to mutate tags and created a comprehensive end-to-end design to support tagging in 18 of Lex's existing build-time APIs involving distributed workflows.
Worked part-time for the Advanced Information Security department. Researched adversarial attacks on machine learning models with a focus on defenses.
Researched deep learning solutions as part of the Mission Algorithms R&S Department: 1. Researched and developed novel methods of pruning convolutional neural networks. Developed a highly efficient image classifier, obtaining state-of-the-art performance on the CIFAR 10 dataset. Placed first in a deep learning competition hosted by Sandia using this classifier. 2. Researched the use of LSTM-based autoencoders with attention mechanisms in creating sequence embeddings (vector representations for sequential data). Methods developed were applied to the domain of source code analysis for cybersecurity; problems explored include supervised classification of source code, generation of source code from pseudocode and unsupervised clustering of source code. The work was presented at Sandia's 2018 ML/DL conference. 3. Gave a paper presentation on Generative Adversarial Networks as part of Math Symposium held in the department.
USC's Leading Machine Learning undergraduate student organization focused on using Artificial Intelligence for social good. Website: http://caisplusplus.usc.edu 1. Taught CAIS++ developed deep learning curriculum to members. Link: http://caisplusplus.usc.edu/curriculum 2. Developed presentations and gave lectures on topics including linear/logistic regression, neural networks, CNNs and RNNs/LSTMs. 3. Helped members during in-person coding workshops for the above topics. 4. Held office hours to provide additional support to members.
1. Part of undergraduate staff for the class CS 103: Introduction to Programming. 2. Helped students and staff by mentoring lab sessions, holding office hours and grading assignments and exams. 3. Topics covered include C++, pointers and memory management, recursion, DFS, BFS and basic data structures such as linked lists, queues, and deques.