Data Science ’28 @ Chapman University · Break Through Tech AI Fellow · Community Staff @ OpenClaw
I build human-centered AI products and developer tools that make complex systems easier to understand, direct, and trust.
My focus is AI product engineering at the intersection of HCI, accessibility, developer experience, and reliable agent systems. I care about making evidence, system state, privacy boundaries, and failure behavior visible to the people using AI.
These projects show how I turn that focus into learning tools, AI applications, and agent interfaces.
An educational reading companion connecting mathematical papers, Lean formalizations, and focused AI explanations.
Readers can select a sentence, equation, or Lean declaration and request an explanation without losing their place in the proof. The interface keeps four kinds of evidence distinct:
- What the paper states
- What Lean formally verifies
- Relevant prerequisite knowledge
- What the AI generates
Built with: TypeScript, Next.js, Lean 4, PDF.js, Playwright, and the OpenAI Responses API
🚧 Tested release candidate; preparing the first public release.
🏆 Entertainment Track Runner-Up — PantherHacks 2026
A full-stack science-fiction mission simulator where human players and AI-controlled crew members coordinate under pressure.
An AI mission director responds to player decisions while the application maintains structured world state, crew dynamics, tactical handoffs, mission mechanics, and persistent save data.
Built with: React, Vite, Node.js, Express, PostgreSQL, the OpenAI Responses API, and the Anthropic Messages API
A cross-platform OpenClaw plugin that represents local and remote agent activity as animated desktop pets.
The plugin uses native overlays on macOS and Windows while enforcing a strict privacy boundary: prompts, credentials, tool arguments, model output, and internal errors never cross the display bridge.
Built with: TypeScript, Swift, C#, WebView2, and OpenClaw
I make reusable skills and developer tools for agents, especially where technical reasoning, source boundaries, and human control matter.
- Lean 4 Codex Skills: Agent skills for theorem proving and autoformalization, including mentoring, declaration drafting, proof repair, review, build-checked checkpoints, and toolchain diagnosis.
- ClawHub Skill Pack: Published OpenClaw skills for reflection weaving, strategic reading, evidence-first review, and citation provenance.
- 2D Hartree–Fock Lab: A browser-based numerical methods workbench that makes convergence, precision, and rejected unconverged steps visible.
- Codex Native Selector: A Windows customization that gives Codex Desktop a clearer native model-selection interface and portable-copy workflow.
As an independent open-source contributor, I focus on developer experience, safer content workflows, and interfaces that adapt to how people actually work.
- Made uploaded work usable without leaving the conversation through safe previews for code, documents, PDFs, and HTML, followed by bounded spreadsheet and slide-deck previews.
- Improved navigation and accessibility with independently collapsible sidebar sections and pointer, keyboard, and touch controls for channel ordering.
- Let people shape the conversation layout around their preferences with independent alignment controls for their own and others’ messages.
- Added matched search context and search results in a sidebar.
- Added responsive Markdown tables, message reactions, and message editing.
- Added channel mention notifications and pinned messages.
- Fixed right-aligned Markdown rendering.
- Added resolved mention highlighting and named responding-agent status across channels and threads.
- Restored reliable end-to-end media delivery for ClickClack, made retries safe against duplicate messages, and removed a hard-coded model limit so replies follow the selected runtime and model budget.
- Added opt-in ClickClack group-reply mention gating, preserving direct-message behavior while preventing ordinary group traffic from activating every account.
- Preserved original cron timeout and cancellation reasons through prompt handoff and finalization so operators can diagnose failed runs accurately.
- Fixed restart recovery for deleted or unconfigured agent stores while keeping configured custom stores eligible.
- Fixed delegated expansion reply collection for current OpenClaw beta releases by using the supported session-message API and adding regression coverage.
- Human-centered AI: HCI, accessibility, evidence-aware interfaces, and AI for technical learning
- Agent systems: Observable workflows, bounded data flows, privacy boundaries, and failure behavior
- Formal and scientific computing: Lean 4, mechanized reasoning, numerical methods, and browser-based research tools
- Stack: TypeScript, React, Next.js, Node.js, PostgreSQL, Playwright, Swift, C#/.NET, Rust/WASM, OpenAI and Anthropic APIs
- Building Interactive Proof toward its first public release
- Developing 2D Hartree–Fock Lab as a browser-based numerical methods workbench
- Creating reusable skills and tools for agent workflows
- Studying Data Science at Chapman University
- Participating in the Break Through Tech AI program
- Building and contributing to open-source AI tooling