ProfileSix years of building, from perception for driverless cars at university to load testing at cloud scale. Here is the long version.
I am a software engineer at GitHub, part of Microsoft, working on GitHub Pages. Until recently I spent four years on Azure Load Testing (opens in a new tab) in the Cloud + AI division, where most of my time went on the service side of things: making large test runs start quickly, run reliably, report honestly, and fail in ways that tell you something useful.
Before that I read Computer Science at BITS Pilani, where I spent three years in Sally Robotics, a student-run driverless vehicle group that met outside class hours. I worked on the perception side, and that work became two papers: one on scene parsing for roads without lane markings, and one on restoring images shot through haze at night.
Away from work I like problems that end with a machine at home doing something it was not really meant to do, usually over a network connection that should not have worked.
Working on GitHub Pages.
Backend for Azure Load Testing in the Cloud + AI division.
Joined the Azure Load Testing team full time after interning with it, and was promoted to Software Engineer 2 in 2024.
A student-run driverless vehicle group on campus, built around whatever we could get done outside class hours. I worked on perception: semantic segmentation and object detection on Indian road scenes.
Interned with the Azure Load Testing team in the Cloud + AI division.
Experimental features using Cisco IOx Docker on Catalyst series wireless LAN controllers.
Automated the analysis of client data using Python, NumPy, Pandas and MLxtend.
Built a predictive text typing system for email clients, along the lines of keyboard auto-suggestion, using PyTorch, Hugging Face and NLTK.
A GNOME Shell extension that puts Cloudflare 1.1.1.1 WARP in the top panel, so switching the tunnel on and off no longer means dropping to warp-cli. Published on the GNOME Extensions store.
A Double-DQN reinforcement learning agent trained and evaluated on the Arcade Learning Environment, built with PyTorch, Gym[atari] and NumPy.
A modified DeepLab V3+ architecture for semantic segmentation of Indian road scenes. Placed in the top 16 of the AutoNUE deep learning challenge.
Object detection on a custom dataset of Indian speed breakers, comparing YOLO V3 against TridentNet.
A computer vision approach to nighttime dehaze-enhancement, restoring scenes degraded by both haze and low light in a single end-to-end model.
Scene parsing for autonomous navigation where lane markings, kerbs and road boundaries cannot be relied on, as is typical of Indian roads.