Senior Full-StackUX Engineer

Merchant onboarding ate a full cross-functional team. Five days of meetings per partner to gather configurations from sales, ops, risk, and the partner themselves. Six entity types, each with about ten metadata fields, arrived through email and Slack threads. Tickets dropped, configs drifted, and stakeholders had no visibility into program status without booking a meeting. I designed OnboardIQ from a blank canvas: a flexible workflow product that codified the entire process into configurable, partner-aware steps. Audits and automation went directly into the workflow, so validation, status, and handoffs were surfaced inline instead of chased over Slack.

Lab technicians couldn't trust AI answers without seeing the source documents — existing tools gave answers without verification. I built a split-screen verification interface with citation deep-linking to exact PDF coordinates so users validate every response. Designed a gap-detection UX with an 85% confidence threshold that asks for missing documents instead of guessing. The pattern was later extracted into Document Domain Agents, a domain-agnostic framework for high-stakes RAG.

Most RAG systems guess when they don't have enough information, creating dangerous false confidence in safety-critical contexts. Document Domain Agents is the framework I extracted from Rayni for any domain where AI accuracy is non-negotiable. The gap-detection UX turns AI limitations into collaborative moments — users upload missing documents instead of losing trust in the system. Verification is split-screen, BLUF formatting handles scannability, and confidence indicators are calibrated to domain risk.

Qualified is a Chrome extension that lives in the browser side panel while you shop. Tell it what you want — "a laptop under $1,000" — and it searches products, compares them side-by-side, and surfaces real financing options from three lenders the moment intent shows, not at checkout. Architecture: React side panel → Express server on Fly.io → Claude (sonnet-4-6) with eight tool definitions → tool execution → streamed back. The wedge is MCP Apps: an extension of Anthropic's Model Context Protocol where tools return interactive HTML UIs, not just text. Sandboxed iframes hydrate via postMessage to render comparison tables, payment calculators, product grids, and pre-filled application forms directly inside the conversation. Same multi-lender waterfall pattern as Versatile / Synchrony — reimagined as an agentic sidecar with generative UI.

MindPattern is two systems wired together. The backend is a deterministic 12-phase Python pipeline that operates as a one-person media company on autopilot: preflight data collection across eight sources (RSS, Hacker News, arXiv, GitHub, Reddit, Twitter, YouTube, LinkedIn), parallel dispatch of thirteen specialist research agents, synthesis of 150+ findings, newsletter publishing, and platform-native social posting. A self-improving harness finds bugs in the pipeline, writes fixes using TDD, reviews its own PRs, and merges them. The public site is a chat interface built around generative UI — when the AI invokes MCP tools that return structured data, React components (finding cards, source tables, health dashboards, pattern lists, skill cards) render directly inside the conversation instead of plain text. Wrapped in a Wire Room intelligence-dossier aesthetic: JetBrains Mono everywhere, manila palette, stamp badges, grid-paper textures.