When AWS set its mandate to embed generative AI across every service, the OpenSearch Observability team had to reimagine DevOps workflows through the lens of GenAI. I joined as Lead Product Designer and, over the life of the project, grew into the team's AI Design System Governance lead—owning both the product experience and the system that let it scale.
OpenSearch is no small canvas: an open-source, Linux Foundation analytics platform, a top-4 search engine, with 750M+ downloads. Building on a year designing AI-driven search-comparison tools, I led the discovery phase—mapping the intersection of machine-learning capability and practitioner need through strategic JTBD frameworks.
Working inside the Nielsen Norman design-thinking frame—Understand, Explore, Materialize—I anchored the team on a single canonical workflow: root cause analysis, the 20% of the product driving 80% of the experience.
The journeyI mapped how an incident actually unfolds: a watch buzz (“Can this wait? It can't.”), a glance on the phone (“How bad is it?”), then to the desktop where triage begins—dashboard, service map, traces, log analysis.
The four steps to root cause lived in four disconnected corners of the product. The path was there; the product made you assemble it yourself.
What practitioners told us — primary research“I feel like I need a PhD to fix slow queries.”
“I get three options and none work how I expect.”
“Cryptic errors, useless autocompletes, buried menus.”
Sourced from OpenSearchCon, 1:1 interviews, and solutions architects.
Setting table stakes — secondary researchI benchmarked conversational AI across Claude, Copilot, ChatGPT, Gemini, Galaxy AI, and Apple Intelligence. Three patterns set the floor: real-time object generation, multi-turn conversation, and stage-appropriate suggested actions.
To align 30 people fast, I ran a hybrid See / Think / Feel / Do workshop—fusing Amazon's Working Backwards method with persona empathy mapping. Two days, one hour each, oriented to a single North Star. Three findings emerged.
“White-glove. Walks you through your workflow as you ask questions.”
“…understand relevant signals so they can quickly find the right problem to solve.”
17 candidate workflows → one undisputed direction → 60–70% of use cases.
Democratize access · Approachable by design · Data-aware.
I moved from wireframes—where product defined the AI narrative and engineering validated stateful interactions—to atoms (AI controls, suggestions, response patterns, input fields) governed by clear heuristics: semantic coherence, economy of form, signal-to-noise, interaction cost. Those atoms composed up into full multi-turn panel organisms.
One decision split the org: where should the assistant panel live—left, right, bottom-docked, or default fullscreen? Rather than let opinion decide, I settled it with research.
The studyUnanimous. Docking flexibility stayed on the roadmap—prescribing a single rigid layout would have strained open-source community expectations.
History, editable names, discardable threads—plus a hidden gem: saving conversations as notebooks to train teammates on RCA. Shipped at launch.
Exposing chain-of-thought taught users the query language. The assistant quietly became an onboarding tool.
Two workflows were ready. Over the flashier alert demo—rich object generation, heat maps—I championed the end-to-end journey, troubleshooting 500 errors, for its context awareness and concierge walk-through. It shipped.
The OpenSearch Assistant Toolkit debuted at re:Invent AWS and OpenSearchCon India. I designed the end-to-end experience for its flagship capabilities—conversational assistants, natural-language-to-visualization, AI-powered anomaly detection, NLQ summarization, and agentic reporting—setting a new benchmark for intelligent observability in the DevOps ecosystem.
On one of the most technical platforms in the AWS portfolio, usability climbed from 'poor' to the top decile.
Shipping the product was half the story. The harder problem: keeping a fast-moving GenAI surface compliant across a team I didn't directly manage. My answer was to stop treating governance as meetings and start treating it as infrastructure.
Videos, screenshots, and text records for every user flow.
GenAI specifications with redlines linked back to Figma canvases.
UX–engineering severity ratings, cross-referenced across all flows.
The endgame proved the thesis. I built a standalone multi-turn tool that generated design-system-compliant Figma templates from a plain-language query—in 2–3 minutes.
Omnichannel and multi-brand consistency, encoded as rules a machine can follow—a human- and machine-readable automation system, and the new norm for AI-ready design.
The North Star workshop is why 30 people across four time zones didn't miss a three-month deadline.
The same rigor that kept the panel compliant is what let AI agents generate compliant design later.
Twice-weekly reviews turned pressure into momentum.
Amazon Ads
Paychex, Inc.
Amazon
Paychex, Inc.
