09:00 · Atlanta

By day, I run product for IT automation.

I’m Xuyang Zhang, Product Line Manager at Omnissa.

10:00Career

Four company names on my résumé. I never left.

I started as an escalation engineer, moved into consulting, then into product. Twelve years in B2B SaaS, helping customers get value from the product. Now I own the roadmap for Freestyle Orchestrator.

  1. LV 12015 · AirWatch by VMwareEscalation engineer
  2. LV 22018 · VMwareConsultant, senior consultant, then consulting architect
  3. LV 32023 · VMware, then BroadcomSenior product manager
  4. LV 42024 · OmnissaProduct line manager
  • 500+ support requests a month, deflected
  • 30+ enterprise rollouts
  • Side quest: PM rotation, 2022
  • +131% workflow runs, year over year
  • AirWatch by VMware
  • VMware
  • Broadcom
  • Omnissa
18:40

By night, I build with AI agents.

Agents get real work when the risky decisions stay outside the model.

19:30 · After hours

Five projects, all built with AI coding agents.

01 / 0520:00Public · v0.1

Agent tools for Workspace ONE

An AI agent runs device management. A person approves the dangerous parts.

A personal project on the public Workspace ONE API. Not an Omnissa product.

Agent · ws1 CLI
  1. ws1 mdmv1 devices search --user alice@example.com
  2. {"ok":true,"count":2,"devices":["Alice's iPhone 15","Alice's MacBook Pro"]}
  3. ws1 mdmv2 commandsv2 execute --commandName Lock
  4. {"ok":true,"status":"Queued"}
  5. ws1 mdmv2 commandsv2 execute --commandName DeviceWipe
  6. Approval required (1 target). Waiting up to 5m0s ...
  7. {"ok":true,"approval_request_id":"req_k3x9p2"}
  8. ws1 audit verify
  9. {"ok":true,"total":2,"failures":null}
A person · 127.0.0.1 · one-time page

DestructiveReversibility: none

DeviceWipe

mdmv2.commandsv2.execute
Alice’s iPhone 15alice@example.com · Apple · Enrolled
ApproveDeny
Approved
Opens in a person’s browser. The agent can’t see it or click it.
  • 980operations
  • 83need a person’s approval
  • 50device cap on bulk writes without approval
How it works

A command-line tool and Claude skill that let an AI agent do routine work in Workspace ONE, the console IT teams use to manage company laptops and phones.

The hard part. An agent that can wipe a laptop must never do it alone. Destructive commands open a one-time approval page the agent can’t see or click, and every change lands in a tamper-evident log.

Built with. A Go CLI generated from 10 OpenAPI specs, shipped with a Claude skill. Anything missing from the policy file counts as destructive. An earlier MCP server with 21 tools is public too.

02 / 0521:00TestFlight

One of Them Is Lying

A detective game where every suspect is an AI on your phone.

Pick a case
Meet the suspects
Question them in your own words
Accuse the liar
  • 56cases
  • 2.8 sper answer, on the phone
How it works

Question suspects in your own words, find the contradiction, and confront the one who’s lying. The AI runs on the iPhone itself, with no server.

The hard part. Small on-device models make poor liars: they drop the story or never confess. Each suspect now holds a written cover story, and the game decides when the right evidence breaks it.

Built with. SwiftUI on iOS 26 with Apple Foundation Models: one model session per suspect, with voice in and out. A content linter keeps cases separate from the engine, and a test suite plays every case’s solution against the real model.

03 / 0522:00Beta

Aloud

Dictation that never sends your voice anywhere.

fnHold to talk · release to insert

You say

So um, can we move the — uh, no, push the review to Thursday and send Maya the deck

  • Off
  • Professional
  • Concise
  • Friendly
  • Bullets

What lands

Could we move the review to Thursday and send Maya the deck?

  • On your Mac
  • No account
  • No cloud
  • No telemetry
  • 1key to hold
  • 3on-device speech engines
Launch video · 1:22
How it works

Hold one key, speak, and clean text appears at your cursor in any Mac app. Recognition and cleanup both run on the Mac.

The hard part. Private has to be provable. Automated tests check for any outgoing network traffic, and IT can switch dictation history off across a fleet but can’t switch it on.

Built with. A Swift 6 menu-bar app. Parakeet, Parakeet Realtime and Whisper run on Core ML behind one protocol; cleanup uses Apple Foundation Models with a local Qwen3 fallback. 1,546 offline tests.

04 / 0523:00Alpha

htmd

Agents build interactive pages for about 95% fewer tokens, and get every edit back as one prompt.

decision-matrix.yaml
question: Which database should we use for the v2 backend?
criteria:
  - name: Query performance
    weight: 5
  - name: Operational simplicity
    weight: 4
  - name: Cost (monthly $/k req)

Edit a score → Copy all changes → one prompt back

  • 24 templates
  • Decision matrix
  • Kanban board
  • Approval list
  • Slide deck
  • Code review
  • PR review
  • Code diff
  • Dashboard
  • Status report
  • Timeline
  • Data table
  • Chart
  • Checklist
  • Form
  • Rank order
  • Priority matrix
  • 3-up comparison
  • Concept explainer
  • Q&A
  • Text redline
  • Feedback corrector
  • Prompt tuner
  • Email digest
  • Network graph
  • ~95%fewer tokens
  • 24page templates
  • 0extra model calls
How it works

Turns an AI agent’s answer into an interactive page. There are 24 templates: a decision matrix, a kanban board, an approval list, a slide deck, a code review, a dashboard and more. The reader edits the page and sends the changes back as one prompt.

The hard part. An agent writing a whole web page is slow and costly. In htmd the agent writes a few lines of YAML and a template draws the page, with dark mode, print styles and drag and drop built in. The decision matrix here is about 360 tokens of YAML; the finished page is 9,000 to 12,000 tokens of HTML.

Built with. A Node 20 CLI and library that writes one self-contained HTML file, with an optional serve mode. MIT licensed.

05 / 0500:00Live

Content engine

One loose idea in. A finished YouTube explainer out.

Loose idea

Why do long AI chats forget what you told them?

IngestResearchBrief
I pick the style · 15 built in
  1. ScriptI mark it up
  2. Voice
  3. Shot plan
  4. Visuals
  5. Music
  6. Presenter
  7. Captions
  8. Assemble
  9. Metadata
  10. Render checkI review it
  11. Publish kit
I publish
  • 31videos published
  • 11stages per video
  • 15video styles
How it works

A production line for an AI explainer channel. It researches, writes, narrates, illustrates, captions and assembles each video. I pick the style, mark up the script, review the cut and publish.

The hard part. Early scripts sounded staged, with the narrator asking itself questions. I rebuilt the script stage around 3Blue1Brown’s teaching sequence, and a checker now flags those tells.

Built with. A Next.js console over a SQLite job queue and a background render worker. Claude writes every language step, GPT Image draws, narration runs on a local voice model, and Whisper and ffmpeg assemble each cut.

01:00 · Also after hours

Eight more.

  • A / 08

    AlgoTrade

    A trading system that runs itself every market day: futures, stock and options strategies on a paper account, each tested three ways before it joins.

    Paper trading
  • B / 08

    Skill Atlas

    A game-style map of 138 AI-builder skills that your own coding agent fills in. The night sky on this page comes from it.

    Prototype
  • C / 08

    Website toolkit

    449 components and a scaffold for small-business sites, shown in eleven demo sites that each start from a different structure.

    Demo
  • D / 08

    OneTake

    A Mac screen recorder that edits its own demos: auto-zoom, captions, AI narration and a timeline editor.

    Beta
  • E / 08

    Kneady Cats

    A cozy tile-matching game for iPhone. Every cleared board rescues a clay cat, and a solver checks each board can be cleared.

    TestFlight
  • F / 08

    LittleSpoon

    A shared log for two parents starting a baby on solid foods: foods tried, allergens introduced, how each meal went.

    TestFlight
  • G / 08

    Storybook Studio

    Turns a toddler’s few words into an illustrated, narrated picture book on an iPad.

    Family use
  • H / 08

    What to Eat

    Settles a household’s nightly “what should we eat?” with a wheel, a slot machine and a week planner.

    Family use
02:00 · Skills

46 skills I can ship.

Every star is a skill I can ship, used in work other people rely on.

Building8
Programming fluencyAPIs and datacontractsVersion controlEnvironments anddependenciesAsynchronous systemsApplicationpersistenceDebugging and testingAI-assisteddevelopment
Product & experience8
User and problemdiscoveryAI opportunityassessmentPrototyping andvalidationInteraction designUser control andrecoveryOutcomes and adoptionFeasibility andtradeoffsDomain and stakeholdercollaboration
Voice5
Audio fundamentalsSpeech-to-text (STT)Text-to-speech (TTS)Streaming audioSpeech activitydetection
Evaluation5
Success criteriaEvaluation datasetsAutomated checksModel-based evaluationRegression evaluation
Prompting & context4
Task instructionsStructured outputsContext assemblyConversation state
Tools & workflows4
Tool interface designService integrationWorkflow orchestrationHuman review steps
Security & privacy4
Identity andauthorizationTool permissions andsandboxingPrivacy and retentionAuditability andgovernance
Model literacy3
Generation controlsHosted and localmodelsAI limitations anduncertainty
Shipping & operations2
Application deploymentCaching and batching
Agents1
Agent suitability
Vision & media1
Image generation
Model adaptation1
Model serving
74 more I can apply

Building

Visual and low-code building

Product & experience

Trust and uncertainty communication · Accessible AI experiences

Voice

Turn-taking and interruption · Speaker and language handling

Evaluation

Human evaluation · Grounding and factuality · Agent evaluation · Robustness and adversarial testing · Online improvement

Prompting & context

Examples and demonstrations · Context budgeting · Task decomposition · Prompt experimentation · Output recovery

Tools & workflows

Tool invocation · Integration protocols · Durable execution · Safe retries and side effects

Security & privacy

Threat modeling · Secrets management · Output validation · Content and misuse controls · Bias and inclusion

Model literacy

AI task framing · LLM mental models · Tokens and context limits · Model selection · Embedding concepts · Quantitative reasoning · Predictive ML basics

Shipping & operations

Release pipelines · Configuration and versioning · Observability · Latency optimization · Cost and unit economics · Load and capacity management · Resilience and fallbacks · Safe rollout and rollback · Incident response · Feedback operations

Agents

Agent execution loops · Planning and replanning · Memory systems · Delegation · Shared state and coordination · Execution environments · Browser and computer interaction · Execution budgets · Agent recovery · Action verification

Vision & media

Image understanding · OCR and layout · Visual detection and localization · Image editing and consistency · Video understanding · Video and audio generation · Multimodal context · Media pipeline design

Model adaptation

Adaptation decisions · Distillation and synthetic data · Quantization and efficiency

Data

Source selection and provenance · Data ingestion · Document parsing · Data cleaning · Metadata design · Dataset curation · Data lifecycle

Knowledge & retrieval

Lexical search · Embedding generation · Grounded answer generation · Structured and graph retrieval · Retrieval failure handling

Self-assessed on Skill Atlas, a 138-skill map for AI builders. Can ship: used in work other people rely on, failure cases handled, maintained. Can apply: shown on a bounded, realistic task.

06:00

Let’s talk.

Have a question about a project? I’m happy to talk it through.