What I work on

The goal is more useful work out of the compute Microsoft already has, so that more cores are available for customers and Microsoft's own services run more efficiently. That includes right-sizing workloads, choosing suitable VM sizes, and provisioning Kubernetes nodes as demand changes, including Karpenter-based node auto-provisioning.

The work spans Azure Kubernetes Service and Virtual Machine Scale Sets. I built the measurement model used to assess whether efficiency work is paying off and help set product direction across the teams involved. The fleet this applies to is large, and the work is cross-organizational: engineering, capacity, support, and finance all have a stake.

I own agentic operations for Kubernetes clusters on AKS: what an agent should be able to investigate, what it may change, how to measure whether it is any good, and how that knowledge ships. The visible piece so far is AKS Skills.

I led the public launch of Azure/AKS-Skills, an open-source collection that helps AI coding agents work with Azure Kubernetes Service. It covers tasks such as troubleshooting, cost optimization, and GPU operations.

The approach is investigate first, then ask before making changes. The skills use the open Agent Skills standard so that useful operational knowledge can travel across compatible agent tools, including GitHub Copilot, Claude Code, and Azure SRE Agent. It launched on August 7, 2026 with seven skills and passed 100 weekly active users.

On October 8, 2026, my pull request merged the operational skills into GitHub Copilot for Azure. They now ship as the official aks-skills plugin, an optional add-on to the base Azure plugin, which reaches far more people. The plugin covers live troubleshooting, known-issue lookup, GPU inference, and packet capture. The base plugin keeps baseline AKS help and offers the add-on, with consent, when deeper investigation would help. Install it in Copilot CLI with /plugin install aks-skills@azure-skills. The original Azure/AKS-Skills repository is archived.

My part was the product side: the case for a dedicated repository, the read-only-first design rule, the launch, the benchmarks, the evaluation approach, and the move into GitHub Copilot for Azure. I also write code for it; my merged pull requests are public. The skills themselves are the work of a team.

Chaos engineering deliberately introduces failures so teams can find weaknesses before a real incident does. I have been the product manager for Azure Chaos Studio since 2023: adoption, enterprise capabilities, and the product direction behind a ground-up platform rebuild. I still own it, alongside the AKS work.

That rebuild is now public. On July 1, 2026, Azure CTO Mark Russinovich announced Chaos Studio Workspaces in public preview. Instead of assembling faults one at a time, you point a Workspace at a subscription or resource group, it discovers what you have deployed, and it recommends named scenarios that mirror real outages: an availability zone going down, a DNS outage, a Microsoft Entra ID outage, a database failover under load, a cache stampede, messaging disruption. Each run produces a report you can hand to a compliance review or a post-incident retrospective. There is also a Chaos Studio skill for GitHub Copilot and an MCP server, so an agent can run the same tests. General availability is targeted for late 2026.

Ashley and I recorded the launch demo for the Microsoft Azure channel. It goes from an empty Workspace to a finished run report, then shows the same tests from the CLI and a CI pipeline.

play 8:42

I wrote most of the public documentation for it. The Workspaces overview, the scenario catalog, and the guide to testing AKS workloads are mine, along with about thirty other pages on Microsoft Learn going back to 2023.

I demoed the product on stage at Microsoft Ignite 2023 and Build 2024. On the same team, I pushed for agent-assisted development, and the team's feature throughput rose roughly 30x in a year.

I started at Microsoft in 2021 on Windows, working on recovery and efficiency. My projects included Windows Optional Features, operating-system recovery over Wi-Fi before Windows starts, and manufacturing tools that helped OEMs build Windows-on-Arm devices using familiar imaging pipelines.

The Wi-Fi recovery work meant getting drivers into UEFI firmware across four silicon vendors. It now ships on major OEM devices. The Windows-on-Arm work was with HP, ASUS, Acer, and Lenovo, so they could reuse the imaging pipelines they already had for x64.

US patent: Automated Resiliency Validation

Granted US patent · Microsoft Patent Award · January 2026

Microsoft gives a Patent Award when a patent is granted. Mine is a black cube with a gold medallion carrying my name and the name of the invention.

The idea came from the Chaos Studio years. Chaos engineering is good at breaking things on purpose. The part that usually gets hand-waved is what happens next: someone watches a dashboard, and if nothing obviously catches fire it counts as a pass. Automated Resiliency Validation does that check automatically, so the system has to prove it recovered. I'm a named inventor. The patent number, grant date, and full inventor list aren't published here.

Microsoft Global Hackathon 2025, first place

1st place · executive challenge · judged under Scott Guthrie

I won first place in the executive challenge at the Microsoft Global Hackathon 2025, judged under Azure CEO Scott Guthrie. The executive challenge is one category of the hackathon. Our team built an MCP-based scanner that looks at AI-generated apps and finds the places they will fall over.

It is the patent's question pointed at a newer problem. The project went on to inspire the Azure Copilot Resiliency Agent, now in preview in the Azure portal, which helps teams design, assess, and fix the resiliency of their Azure workloads.

On the side

I have built Agency Co-Work, a desktop workspace for AI agents, and a pediatric-dosing iOS app used by Doctors Without Borders teams. I also built the backend for the AI on this website and run its gateway on a Mac mini at home.

I studied computer science and philosophy at Duke and spent a semester studying philosophy at Oxford. I grew up between the US and India and now live in Seattle.

Outside work: films, Duke basketball, a scratch map with plenty left to do, LEGO that takes over the shelves, and a garden that produces too much basil.

If something here overlaps with what you're building, email me at nkaul2345@gmail.com. The resume has the dates, and my AI can answer follow-up questions.