Enhanced and integrated the Cases at Risk model into Microsoft’s DfM workflows, driving $20M in annualized cost savings by proactively identifying and addressing high-risk support tickets.
Led the development of a high-quality golden dataset and optimized vector database architecture for Copilot for Azure, significantly improving retrieval accuracy and response quality for enterprise customers.
Architected and deployed a comprehensive experimentation platform supporting causal, Bayesian, and frequentist evaluation methodologies across 10+ product teams, driving over $100M in impact through improved decision-making and optimization.
Conducted a comprehensive measurement study of Support Copilot’s impact on agent productivity and customer satisfaction, establishing key metrics and success criteria for AI-powered support tools.
Contributed TabTransformer architecture to the AutoGluon open-source project, enhancing its capabilities for tabular data processing and improving model performance on structured datasets.
Published in Stanford CS238: Decision Making under Uncertainty, 2019
Abstract: We present a stereo-camera-based 3D vehicle-tracking system that utilizes Kalman filtering to improve robustness. The objective of our system is to accurately predict locations and orientations of vehicles from stereo camera data. It consists of three modules: a 2D object detection network, 3D position extraction, and 3D object correlation/smoothing. The system approaches the 3D localization performance of LIDAR and significantly outperforms the state-of-the-art monocular vehicle tracking systems. The addition of Kalman filtering increases our system’s robustness to missed detections, and improves the recall of our detector. Kalman filtering improves the MAP score of 3D localization for moderately difficult vehicles by 7.7%, compared to our unfiltered baseline. Our system predicts the correct orientation of vehicles with 78% accuracy.
Download here
Published in Stanford CS224U: Natural Language Understanding, 2020
Abstract: Child Speech Recognition (CSR) is a less explored and more challenging task than typical Automatic Speech Recognition (ASR). This task has significant applications in the classroom and is especially important in a remote learning environment. We present findings from training deep-learning based speech recognition models on the MyST corpus, the largest publicly-available English language child speech corpus. We obtained 27.26% word error rate (WER) on the MyST test set with a DeepSpeech2 baseline. Our best model, a Conformer model pre-trained on LibriSpeech and fine-tuned using the MyST corpus, achieved a test WER of 23.45%. Our results show that pre-training on adult speech is essential for model performance. We also provide additional error analysis on our best model and discussion of the results.
Download here
Published:
This is a description of your talk, which is a markdown files that can be all markdown-ified like any other post. Yay markdown!
Published:
This is a description of your conference proceedings talk, note the different field in type. You can put anything in this field.
Undergraduate course, Stanford University, Computer Science, 2019
Course Assistant for CS 103: Mathematical Foundations of Computing.