# Haochuan Zhang — full public portfolio text The complete public About page and project index follow, with source URLs for attribution. This is a static snapshot, not a live service or a release-status report. Journal articles are not bundled; discover them through https://blog.haochuanz.net/llms.txt. The browsing guide is at https://blog.haochuanz.net/skill.md. The source bodies retain their original Markdown and links. For a root-relative link beginning with /, append that path to this deployment root (without its trailing slash): https://blog.haochuanz.net/ --- # About Haochuan Source: https://blog.haochuanz.net/about I’m Haochuan Zhang, a software engineer and independent builder working across agentic software engineering, local AI infrastructure, personal computing, and wearable interfaces. Most of these projects return to one practical question: **How can individuals turn abundant machine intelligence into durable capability without surrendering agency, context, or accountability?** I think of the answer as a personal intelligence system: the compute that supplies model capacity, the agents and tools that perform work, the interfaces through which a person directs them, and the permissions and evidence that keep that person responsible. ## What I build My most visible project is [CyberHUD](https://cyberhud.net), an iPhone-hosted programmable HUD for compatible USB-C display glasses. It began because I wanted a useful clock in my glasses. It grew into a focused Desktop with native and bundled web applets, optional stereo presentation, and a bounded creation system called **Netics**. The [month-one build log](/posts/building-cyberhud-month-one) records how each useful feature exposed another product or engineering boundary. Netics lets AI create small Custom Applets inside an explicit product contract. Generated source has to fit declared interfaces and pass native validation before the product accepts it, and generation begins only after an explicit user action. The host retains control of capabilities, permissions, credentials, and storage boundaries. This makes personal software easier to create while keeping the safe path visible and understandable. Alongside CyberHUD, I run local-model and personal-compute experiments. I care about what a model can actually do on hardware I control, where it fails, and which surrounding systems make it usable. My [DeepSeek V4 Flash experiment](/posts/running-deepseek-v4-flash-0731-with-48gb-ram), for example, became a whole-machine systems problem involving GPU topology, host memory, NVMe storage, precision, power limits, and failure containment. I also publish [developer and agent tools](/projects), including integrations for Blender, Sketch, and Xcode. These projects occupy different layers of the same stack: compute supplies capacity, agents turn capacity into execution, product boundaries make that execution useful, and personal interfaces keep the person in control. ## How I work with AI In much of my AI-assisted work, producing a first attempt is no longer the scarce part. Context, judgment, integration, and verification are. I use models as collaborators and bounded operators inside an engineering system. Agents may investigate, implement, test, review, or simplify. Before I delegate execution, I try to define the intent, constraints, permissions, and evidence required for acceptance. During research, design, and debugging, I leave room for the original thesis to change when the evidence points elsewhere. AI can accelerate execution. I remain responsible for direction, integration, physical testing, provider spending, safety, rights, publication, and the consequences of the result. A capable model helps, although the durable multiplier comes from the system around it: specifications, tools, validation, tests, repeatable handoffs, and human review. The model is replaceable. The context, standards, and responsibility should remain mine. I also try to keep the boundaries of evidence explicit. A passing test proves only what it tested. A successful render does not prove physical comfort. A reproducible artifact can still lack approval, suitable rights, or a truthful public claim. When I do not know, I would rather say so and continue investigating. ## What I write The work here tends to fall into several connected areas: - agentic software engineering, orchestration, and human–AI collaboration; - local models, personal compute, and AI infrastructure; - CyberHUD, wearable computing, and focused display interfaces; - personal knowledge systems, memory, permissions, provenance, and governance; - product build logs, validation, release work, and design decisions; - learning, cognitive independence, and human agency; - speculative technology and culture; and - occasional travel, photography, gadgets, and internet weirdness. Some posts report direct implementations, measurements, failures, and reproducible experiments. Others are interpretation or scenario-building. I try to keep those modes visible so a compelling story does not quietly become an established fact. ## Why this blog exists This blog is my long-term memory, build journal, and intellectual test bench. Product pages should describe what currently exists. Repositories should hold reproducible artifacts. This blog records the slower layer: why I made a decision, which paths failed, what evidence changed my mind, how a system evolved, and which claims remain uncertain. Writing in public makes ideas easier to examine and harder to quietly rewrite after the fact. It leaves me a record of what I believed, what I built, what worked, and where I was wrong. I keep that record on a part of the web I control. Social feeds are useful for distribution, but they are poor long-term memory. I want this site to behave more like a notebook than a casino. The original essay, [“Why I Keep a Personal Blog in the Age of AI”](/posts/why-i-keep-a-personal-blog-in-the-age-of-ai), explains that choice in more detail. My current answer to the question behind this site is a build program: make intelligence useful, bounded, inspectable, portable, and answerable to the person using it. This blog is where I keep the evidence, revisions, and parts that did not work. ## Explore next - [Building CyberHUD, Month One](/posts/building-cyberhud-month-one) connects the product, Netics, concurrent agents, and document-driven engineering. - [The Age of Intention](/posts/age-of-intention) explains how I organize bounded agent work while remaining the human integrator. - [I Ran DeepSeek V4 Flash 0731 Locally with 48 GB of RAM](/posts/running-deepseek-v4-flash-0731-with-48gb-ram) follows a local-model experiment from misleading failures to a measured system. - Browse my [public projects](/projects) and their [source on GitHub](https://github.com/lunarmoon26), try selected [CyberHUD applets](/applets), or continue through the [full archive](/archive). ## Colophon This is a statically published site built from Markdown and version-controlled source. I do not run behavioral analytics or engagement mechanics. --- # Public project index Source: https://blog.haochuanz.net/projects Structured catalog: https://blog.haochuanz.net/api/projects.json I keep this page as a living index of the projects I publish on GitHub. I will update it as the work evolves. ## OpenCode plugins and skills Each project combines an OpenCode-native plugin with focused skills for a local or hosted capability. - [OpenCode Blender](https://github.com/lunarmoon26/opencode-blender) — connects OpenCode to Blender's local MCP bridge, with workflows for scene work, Poly Haven assets, and Three.js exports. - [OpenCode Sketch](https://github.com/lunarmoon26/opencode-sketch) — exposes Sketch's local MCP through a native tool and provides an adapter skill for Sketch workflows. - [OpenCode Xcode](https://github.com/lunarmoon26/opencode-xcode) — connects OpenCode to Xcode's local MCP and bundles focused Xcode development skills. - [OpenCode Cloudflare](https://github.com/lunarmoon26/opencode-cloudflare) — exposes Cloudflare's hosted code-mode API MCP as one native OpenCode tool, with a focused skill for its workflows. ## Personal agent skills These portable workflows follow the [Agent Skills](https://agentskills.io/) format and can be installed independently of OpenCode. - [Agent Skills](https://github.com/lunarmoon26/agent-skills) — my personal collection of reusable skills for document-driven development, architecture, research, image processing, and Chinese technical writing. ## Agent tooling and runtimes - [Harness Alchemist](https://github.com/lunarmoon26/harness-alchemist) — a native Rust CLI for scaffolding portable coding-agent plugin projects, with shared skill execution across harnesses ([one-pager](https://blog.haochuanz.net/harness-alchemist/)). - [Agent Skill Runtime](https://github.com/lunarmoon26/agent-skill-runtime) — a shared execution contract for packaged scripts, with CLI commands for discovery, preparation, and bounded JSON-in/JSON-out execution, plus optional MCP, OpenCode, and DeepSeek Harness adapters ([implementation notes](/posts/agent-skill-runtime)). ## Other projects - [PrismDeckJS](https://github.com/lunarmoon26/PrismDeckJS) — browser-native spatial presentations with editable PPTX/ODP import, stereo output, Rapier physics, and standalone HTML export ([live Studio](https://lunarmoon26.github.io/PrismDeckJS/)). - [Article TTS Reader](https://github.com/lunarmoon26/article-tts-reader) — Chrome and Edge extension that extracts an article's main text and narrates it through a user-configured local or hosted TTS endpoint. - [DeepSeek V4 Flash](https://github.com/lunarmoon26/deepseek-v4-flash) — a documented local service deployment for DeepSeek V4 Flash on RTX PRO 6000 Blackwell GPUs. - [OpenClaw MCP Router](https://github.com/lunarmoon26/openclaw-mcp-router) — an OpenClaw plugin that uses semantic search to discover MCP tools at runtime and reduce context overhead. - [Star Digital Employee](https://github.com/lunarmoon26/star-digital-employee) — infrastructure as code for long-lived AI employees with versioned recipes, isolated runtimes, and auditable execution.