Design Engineer, AI products
I build AI products from the foundation up, from the systems and platforms they run on to the design systems teams build with and the data-dense interfaces where humans and agents work side by side. I start from the end experience and design everything underneath to serve it.
I care that what I build is structurally sound and good to use. I work and collaborate across design, product, frontend, backend, data, and AI to get there.
Pattern library
-
AI platform
-
Agent UI
-
Agent knowledge
-
Design system
-
Frontend platform
-
Data viz
-
Tooling
My Work
References Résumé-
JupiterOne In-App Agent As our AI features scaled, they grew sideways, with bespoke SSE endpoints, multiple agents, and isolated features scattered across the app, each taking about two weeks to ship. I wrote the RFC, architecture, and prototype for one AI transport, using the UI itself as the agent's rendering engine, and got it on the roadmap. Existing features collapsed into skills, workflows, and tools that the agent or the user calls through one "intent: make", so the assistant can describe the product, search the docs, write queries, fill out forms, and tour you around the app. My thesis for this project was that the existing UI is the best interface for an agent too. I built the router plugin that makes every route declare what it is and its filters, and 233 routes compile into one typed registry that feeds route guards, navigation, the Cmd-K palette, and the agent's site map. Every data query says what it fetches and the URL holds the state, so the agent sees the page the way you do. Pages publish typed actions that the palette and the agent share, so you both have the same buttons, and the agent opens whatever it's talking about. Drafts wait for your Save and consequential changes wait for your click. It works tucked, as a sidekick docked beside a table, chart, or graph, or full page, with fluid handoff between you and the agent. Assistant 0.7 Personalized home Assistant 0.5 Assistant 0.4
-
Web App Rebuild Our product ran 15 separately deployed micro-frontends on one page, so it showed its seams and couldn't keep its dependencies current, which was a security and maintenance risk. I pitched a replacement to our teams with an RFC, a basic design, and a prototype, got it onto the product roadmap, and led a team of four to build it. I designed the migration from single-spa to one monorepo, its architecture and layers, and worked with a frontend platform engineer and SRE on the new deployables and repo patterns. We built it for AI-first engineering, with governance rules, linting, patterns, and layers set firmly up front. I built the design system on headless Radix primitives, the app shell, TanStack Router with outlet- composed routing, access gates, overrides, preferences, and one chrome layer that gives navigation, the page, and the agent their own surfaces. We moved app by app in vertical slices, mirroring the old look and navigation so customers never felt the switch, then applied the new design language. It was in production in five weeks and active for all users in five months. Along the way we revisited every dependency and replaced a $72,000-a-year visual-testing bill with self-hosted Storybook, visual regression, and unit tests. 82% of surveyed customers now rate the UI "good" or "excellent," where 60% rated the old client "neutral." Since launch, the frontend has caused zero incidents (10 before), deploy failures fell from 4.3% to 0.6%, and I was its largest contributor, with 733 of 2,123 merged PRs. App release notes Search release notes
-
Agent and Intelligence Platform Behind the in-app agent, I rebuilt the agent itself. My RFC treats AI quality as a context-building problem, judged on four questions of correctness, relevancy, intent, and learning. I built one agent loop that can stop and ask instead of guessing, and one tool registry shaped so MCP servers can share it. In two months, 13 feature-specific AI endpoints and a standalone service folded into one agent door. Agents start grounded. Precomputed "hot context" carries what the product, the account, and the user already know, a background prober learns what each customer's data contains, and observed facts feed a knowledge layer with decay, so each answer builds on the last. On a 40-question eval, writing a query went from 36.5 to 17.1 seconds and from 13 to 4 cents a question. The aim is to move the product from pull, where users dig for answers, to push, where it tells them what matters and how it connects.
-
Data Visualization Data viz is critical to great UX. It turns massive amounts of raw rows into something understood at a glance, and it lives in implicit gestures and layers like drill-down, tooltips, date range selection, date handling, and data transformation. Our product runs on a graph database, so I designed and built a new graph canvas on React Flow and elkjs, replacing a commercial graph library licensed at $69,000 a year. We render hundreds of nodes, not millions, so we borrowed from mapping software, using zoom fidelity, collision detection, algorithmic and force-directed layouts, hover dimming, and node packing to turn hairball graphs into pictures people understand. Every behavior helps you get back to where you were, like packing by asset type for spatial memory, dimming instead of hiding, expansions kept in the URL, and a push-pop inspector stack. I started the chart package with D3 for the math and React for the SVG, with a date brush that snaps to real data, trend charts, sparklines that break at gaps instead of inventing values, and one cursor tooltip for every chart. Pointer and drag work writes CSS variables on an animation frame and commits to React once, from the tooltip to the CSS-grid dashboard engine, so dashboards now render 50% faster. On top of the canvas sits Schema Explorer, where you click asset types to build a path through the graph, watch the query write itself, and share a half-built query as a link. Insights release notes
-
Historical Trends Our product traditionally showed only a live picture of a user's data. I built a reporting and metrics service that snapshots live data, so users can answer the question security leaders ask all the time, are we getting better or worse? The graph only knew the present, but three platform machines had been writing daily query results to S3 for up to 8 years with nothing to read them. I designed and built the pipeline, covering capture through the existing query machinery, transcode to Parquet with DuckDB on EKS, rollups, and a report reader serving 41 query templates. That made fast 90-day trends and executive dashboards possible for the first time. Users can inspect many dimensions and individual rows in under 3 seconds, which outperformed our live query service. The agent reads it to say what changed, not just what is. Executive dashboard Vulnerability management 2.0
-
AI Service I was one of four founding engineers of the company's new AI service in October 2025. I defined the SSE streaming contract with real-time rendering, workflows for query construction and validation, the first scored eval dataset, and an LLM-as-judge pipeline. I worked with product to stand up our first eval sets, define failure modes, and set the quality and steering benchmarks that every later AI feature ships against. I also shipped cross-region model routing on Bedrock and later rebuilt the service around one agent loop. Earlier in 2025 I built bring-your- own-model for the compliance AI, with encrypted provider config, admin- only access with audit events, and one AI client interface, so features don't care which model sits behind them.
-
UX Consistency The legacy product felt disjointed. Each vertical slice of the app showed a different era of UX design, which disoriented users. I led a push with four contract engineers to make the experience stable and predictable by solidifying patterns with Design. Six apps moved onto our new virtualized data table in a week, and the rules and alerts tables went from 25 rows per page to about 500, and got faster. I was engineering lead on the navigation redesign. I designed a config-driven, globally mounted inspector that collapsed 12 components into one, simplifying both the experience and development, and shipped dark mode, native internationalization, and WCAG 2.0 compliance. Blog post Release notes
-
Security Products I built across the security products in the legacy app. In 2024 I led the team on exposure management, including an options memo that weighed five ways to compute attack paths and the risk of each. In 2025 came controls monitoring, about 60 PRs in six weeks, plus the framework wizard. Then I split an overgrown repo into three products and deleted about 40k lines. In 2026 I built the vulnerability management backend, with remediation plans grouped in three tiers, AI enrichment behind a circuit breaker, and risk scoring moved from hardcoded values to settings customers can tune. Exposure management launch Framework wizard Vulnerability management 1.0
-
Query Results and Filterable Dashboards Our product's primary use case was to write a query and view the results. But the query results table had a slew of usability issues, with no virtualization, heavy lag when rows had custom renderers, and no column resizing, column reordering, row sorting, row selection, pagination, and more. Tables were hand-built in 12 places around the app. I rebuilt query results on a new data grid, rolled it out to every customer, and deleted v1. Then came my first full-stack feature, dashboard variables, which let a whole dashboard be filtered by safely injecting values into our query language. That meant a new service entity and GraphQL API, a query rewrite that parses each query and drops any filter whose variable is unset (replacing raw string injection), and the drawer and toolbar to set them. The rewrite is chosen by version, so no customer data had to migrate. Drill-through with variable values followed in 2025. New results table Matrix and drill-down
-
Frontend Guild I led the frontend guild. We ran two goals a meeting, cut unique production errors from 51 to 14, set a Core Web Vitals baseline, and made "reproduce it under CPU and network throttling" the rule for slow- page bugs instead of "cannot reproduce." Before leading it, I wrote about 30 team standards in 2023, among them container and presentational components, routable modals, URL as state and links for navigation, then turned the accepted ones into lint rules and shared tooling. I also ran an AI service outage as incident commander. It was caught by end-to-end tests and fixed in about four hours, with three preventive changes.
-
Design System I became the design system's primary maintainer in 2023, with 77 of 95 merged PRs that year. I shipped a new palette and tokens, then led the MUI 4 to 5 migration across 18 repos with old and new styles running side by side, so no app had to switch at once. I designed the typed styling API and component override wrappers, created the navigation package, and later added dark mode that stays in sync across tabs and apps. In 2025 I wrote the case for patterns before components. A product built from a few hardened patterns is faster to build, steadier in production, and easier for AI to understand and extend. It became the plan for the rebuild.
-
Global Search Product had long wanted a search bar at the top of the app to "search anything," but the previous version only understood our query language, and only 3% of visitors used it. I wrote the strategy memo, studied real queries, and modeled the bar on a browser omnibox with history, saved questions, natural language to queries, and full-text fallback. I joined our AI search team the same year and owned the search side of natural-language querying, including feedback, telemetry, retry and the unable-to-answer state. The second generation, a docked bar with a resizable preview drawer, went GA in March 2025, and both later gave way to the command palette. Search bar AI search Preview drawer
-
Asset Inventory I joined in February 2022, and a month in opened the 100-file foundation PR for rebuilding the asset inventory, the product's main index, with three other engineers. I built the filtering that compiles filter chips into queries and the critical-assets logic, then retired v1. Six months in, I created the company's shared table library, with keyboard grid navigation and ARIA from day one, Storybook with visual regression tests, and a render path that took 3,500 cells from 1,000 ms to 350. Critical assets
People I learn from
- Adam Argyle Modern CSS used with taste. Most of my favorite tricks here started with one of his demos.
- Martin Fowler Clear names for architecture ideas, and the habit of writing patterns down.
- Nielsen Norman Group Research-backed usability. The reason I check a design against how people actually behave.
- Edward Tufte Show the data, cut the chart junk. The bar every chart I build gets held to.
- Andrej Karpathy Treat what an AI learns as a wiki it keeps up, not a pile it searches. That idea runs through how I think about agent knowledge.