Product Manager · AI Builder

I build the things I'd otherwise write specs for.

I'm Xuyang Zhang, a product manager on Omnissa's Freestyle Orchestrator team. I use AI tooling to build real, working software — AI agents, developer tools, and privacy-first apps, plus a steady stream of personal and productivity projects. Production-minded, safety-first, and shipped, not slideware.

See selected work GitHub

Selected Work

ws1 Go · open source

An agent-shaped CLI that operates enterprise device management — safely

Lets an AI drive Omnissa Workspace ONE UEM from plain-language goals, without handing it the keys.

Workspace ONE UEM is a sprawling enterprise API. ws1 turns it into something a Claude-class agent can operate from natural language while the human stays in control of blast radius. Every destructive action — wipe, unenroll, delete — and every high-volume bulk op pauses for a browser approval the agent can't fabricate. Unknown operations fail closed. A hash-chained audit log makes every state change verifiable after the fact. The product bet: agents earn a place in the enterprise only when humans own the blast radius.

GoCLI + Claude skillOAuthHuman-in-the-loop

View on GitHub →

workspace-one-uem-mcp Python · open source

An MCP server that makes an enterprise platform programmable by AI

25+ Workspace ONE UEM operations, exposed to any AI assistant through the Model Context Protocol.

Search and command devices, manage users and smart groups, query compliance and audit events — all from a chat interface. OAuth2 with automatic token refresh and bulk operations are built in. Shipped early in the MCP ecosystem as a working bridge between agents and real enterprise infrastructure, not a demo.

PythonMCPOAuth2Enterprise API

View on GitHub →

htmd Node · open source

A Markdown ↔ HTML bridge for AI agents

Agents speak Markdown; humans want HTML. htmd is the bridge — with an export button that closes the loop.

Built on a sharp thesis: an agent writing HTML by hand burns ~2.5× the tokens and skips the polish. With htmd the agent emits ~150 tokens of YAML and gets back a polished, interactive page — 19 widget templates, from live decision matrices to drag-and-drop kanban boards. Every widget can turn a human's edits back into a structured prompt, so review happens in the browser and flows straight back to the agent. Measured at 7–34× cheaper than hand-written HTML.

NodeCLIAgent tooling19 templates

View on GitHub →

aloud Swift · in evaluation

Local, private voice dictation for macOS

Streaming voice-to-text that runs entirely on your Mac. Nothing touches the cloud.

A menu-bar dictation app powered by on-device speech models (Parakeet + Whisper) — no servers, no accounts, no data leaving the machine. Built privacy-first so it's safe to dictate sensitive or work material anywhere. It began as an earlier prototype, VoiceBridge, and was rebuilt from the ground up as aloud — now in evaluation for company-wide use.

SwiftmacOSOn-device MLPrivacy by design

Private · in evaluation

More in the workshop — AI media cataloging, Workspace ONE tooling, and assorted weekend builds live on github.com/xyzbuilds →

In the Workshop

A running list of things I build outside of work — to scratch an itch, chase an idea, or go deep on a personal interest. Some are shipping, some I'm rebuilding with today's models.

personal-finance-trackerRebuilding

Self-hosted finance, categorized by AI

Links US bank and card accounts via Teller, syncs transactions, and uses an LLM (GPT-4 or a local model) to categorize spending. Encrypted at rest and fully self-hosted — your financial data never leaves your server.

FastAPIReactTellerLocal-LLM option

Private · in rebuild

clone-a-clawBuilding

A marketplace for agent personas

Buy, sell, and share AI-agent personas — the bundle of OpenClaw config files (SOUL, IDENTITY, AGENTS, TOOLS…) that gives an agent its character and skills, with the personal files left out. Every upload is scanned for prompt-injection, destructive directives, and leaked secrets, shown as a pass / warn / fail trust badge.

Next.jsSupabaseStripeSafety scanner

Private · cloneaclaw.com

tabagentRebuilding

Tabs that organize themselves

A Chrome extension that de-dupes and groups tabs by the workplace app you're actually using — Microsoft 365, Google, Atlassian, Slack, and 20+ more — via Chrome's native grouping. All local; nothing leaves the browser.

Chrome MV3Vanilla JS25+ app rulesLocal-only

Private · in rebuild

pm-toolkitIn progress

A PM copilot, grounded in my context

PM skills grounded in a "context pack" — Profile, Projects, Counterparts, Vision — so Claude knows my world before it acts (an idea borrowed from how OpenClaw agents load their identity files). Organized around inbound, outbound, and strategy work. I'm user zero; two inbound skills, Daily Triage and Project Pulse, are already part of my daily routine.

Claude skillsContext packsPM workflowsDogfooded

Private · I'm user 0

AlgoTradePersonal

An algorithmic futures trading system

A personal quant project I've gone deep on: I built the data and backtesting harness, found and hardened a strategy, and moved it into paper trading with walk-forward testing — surfaced in a dashboard that flags when to eyeball a manual trade versus let the system run. Python, tested, treated like production software. (Strategy specifics stay private.)

PythonInteractive BrokersBacktestingStreamlit

Private · personal project

How I Build

// 01

Start from the problem

I dig into the real user or business problem before writing a line. Building isn't the point — it's how I pressure-test whether the solution actually holds up.

// 02

Outcomes over output

What did it change once it shipped — adoption, time saved, a clear before-and-after? I measure that, not how much got produced.

// 03

Human-in-the-loop by default

Agents act, people approve. The interesting design work is the approval moment — what to surface, when to pause, how to make consent impossible to fake.

// 04

Trust is a feature

Fail-closed defaults, privacy by design, verifiable audit trails. Trust is what makes powerful tooling adoptable inside a real organization.

// 05

Patterns travel

I lift a good idea from one domain and port it to another. OpenClaw's agent-identity files became the "context pack" that grounds my PM skills in pm-toolkit — and the tradeable artifact behind clone-a-claw. The best primitives don't stay in their lane.

About

I'm a product manager on Omnissa's Freestyle Orchestrator team, focused on how automation and AI agents are reshaping the digital workspace. The fastest way I understand a problem is to build something that solves it — which is why most of what I ship is real, working software, not just specs.

That habit lets me sit between product, design, and engineering and speak all three: framing the bet, sketching the interface, and using AI tooling to build the working version that proves it out. The result is sharper scoping and faster proof of what's actually worth building.

I follow AI tooling closely — often closely enough to lose sleep testing whatever shipped that day, then folding what sticks into my workflow or my next build. I've learned an enormous amount that way, and I'm clear-eyed about how much more there is to learn.

I'm a Certified Scrum Product Owner, previously at VMware, and based in Atlanta.

RolePM, Omnissa

FocusAI agents · dev tools

Builds inGo · Python · Swift · TS

CertCSPO

BasedAtlanta, GA

Let's build something.

Open to conversations about product, AI tooling, and the workspace of the near future.

Email github.com/xyzbuilds LinkedIn