A local plug-in marketplace for my AI coding workflow, with four plug-ins: collaboration rules and hooks, workflow tools, interface design, and code security review.

• Kept in sync between my Mac and a DGX workstation.

A technical report on when an AI agent has really finished a piece of shared work. It proposes evidence-bound completion: every “done” has to point to an artifact that can’t change, the checks that apply to it, and a result the receiving side accepts, with doing, checking, accepting and publishing kept as separate steps.

• Self-published on this site on September 29, 2026.

Read the report

An audio plug-in for mastering music, available as VST3, AU and a standalone app for macOS and Windows.

• A fixed four-stage chain: BALANCE (dynamic EQ), CLAY (stereo cleanup), COLOR (tone and glue) and LIMIT (true-peak limiting).
• A single GLAZE knob listens to the track as it plays and works out how hard to push it.
• It comes with a bilingual website, a user guide and a public page explaining the algorithms, and there is a Windows installer.

celadonaudio.com

I turned the open-source Droppy Code into a native macOS AI coding client for my own use.

• Added a multi-agent mode called Crew, undo that only rolls back files, session logs and project archive pages.
• Installed and in use on my own computer.

A Grok bot posts the day’s briefing to the site at 8 a.m. Each issue gets its own page, with a table of contents, search, topic filters and light and dark themes, in both Chinese and English.

• It runs on my own server; version 0.5.0 launched on September 17, 2026.

brief.maxsong.me

A “shared brain” on my own server. Project documents are kept in Markdown, AI agents read them and claim tasks through a controlled interface (MCP), and people handle sign-in, approvals and editing.

• The read-only workspace and document sync already run on my server; the remaining features are still on a development branch.

Demo (sample data)

I studied the source of OpenAI’s Codex coding agent: how it organizes context, calls tools and decides what to do next. Several AI agents read the code in parallel under my direction, and I checked their conclusions against a pinned version, including the places where their accounts disagreed.

• 77 notes, 11,344 lines in all. The research notes are private.

A native macOS app that scans all of my local AI coding projects and shows each one next to its Obsidian notes, so it’s easy to look up why something was done.

• It went from v0.1 to v0.4 in three days, including a redesign in the Liquid Glass style. It runs only on my own Mac; there is no public download.

A macOS menu-bar app that reads how much Claude and Codex quota is left, suggests a conservative model for the size of the task, and warns the working AI agent through a hook when quota is about to run out.

• Dynamic Island-style alerts and a Control Center widget. Version 1.2.0 shipped on August 11, 2026, and I use it every day.

Demo (sample data)

An AI assistant that runs on my computer, connected to WeChat and scheduled tasks. My customizations: models working together (Grok suggests first, Claude summarizes and acts), 126 skills shared with Claude Code, a separate bot that can operate the computer, and daily work reports written automatically in June and July.

My portfolio, with my projects, technical reports and my own music library. The homepage is a Monet-style painted scene made with Three.js.

• A browser DJ player with a pitch shifter (WSOLA) and a reverb (FDN) that I wrote myself, running in real time in an AudioWorklet.
• Hosted on my home server.

Listen on the Sound page

Semiment makes automotive-grade sensors, driver chips and power chips and has about 300 employees. Over the summer I built systems the whole company now uses, mostly by running several AI coding agents in parallel while I owned the requirements, architecture, task breakdown, review and deployment.

• Three main projects — a company-wide knowledge base, a training and exam platform and the company website — plus an AI gateway for the company (each listed below).
• Three internal tools, all in use: contract stamping by company name or keyword (scanning dropped from about 3 minutes to 3.3 seconds per page), performance reviews inside WeCom, and weekly and monthly sales reports generated automatically.

The company’s bilingual Chinese–English website: the main site, a product documentation site, an admin console and the back end behind them.

• Upgraded every dependency, rewrote the copy and parameter tables for 133 products, and organized 225 datasheets into a product catalog.
• Optimized it for AI search (GEO): 327 prerendered pages, an llms.txt file and rules for AI crawlers, plus AI product Q&A backed by the knowledge base.
• Tracked down and fixed a production incident in which a pop-up locked the whole site. The company signed off on a batch of changes on August 19, 2026, and all of them are live.
• The main work ran from May 25 to August 2026, with occasional fixes since.

A retrieval service that answers questions from the company’s chip datasheets and selection guides. The whole company, about 300 people, now uses it; answers cite their sources, and it declines when the documents don’t cover a question.

• 225 datasheets and selection guides went in; after two problem files were removed, 223 documents became 6,359 searchable chunks.
• Hybrid search combines vector search with BM25 keyword search (with Chinese word segmentation), then re-ranks the results. On the test set it reaches a hit rate of 0.90 and an MRR of 0.85.
• It runs as four always-on services on an in-house GPU machine, with health checks and monitoring across the whole chain, and it also powers the AI chat on the company website, answering product and parameter questions from the datasheets with citations.
• The first phase went live on June 25, 2026. A second phase that could read charts was prototyped, then paused.

The company’s internal training platform, used by every employee; 10–20 exams have been run on it so far.

• An admin console and an employee portal, with WeCom QR-code sign-in and directory sync.
• Exams with online grading and automatic make-up lists, QR check-in with a live attendance board, a course workspace, and reminders through WeCom messages and a mini-program.
• All 913 automated back-end tests pass.
• Built and launched between July 14 and August 19, 2026; since then I’ve been moving it onto the company intranet and maintaining it.

Demo (sample data)

A standalone web app modeled on the AI-gateway feature of 1Panel’s enterprise edition: one OpenAI-compatible API for the company, with model quotas for each employee, API keys and usage audits.

• I wrote the upstream account pool, request forwarding, load balancing and smart routing myself; the router first judges how hard a request is, then picks a model.
• An admin console for desktop and an employee portal for phones, both with WeCom sign-in.
• Built in ten phases; all 3,560 automated tests pass. It has not officially launched yet.

Demo (sample data)

For AP Physics 1, I turned five core concepts into an interactive lab site with adjustable parameters and real-time simulations, with formulas typeset in KaTeX. I wrote a product spec first, then built it.

Demo

Teachers have invited me to DJ four school events. One, in May 2026, was a student original-music showcase where people shared songs they had written or performed well-known pieces; I DJed for about five hours.

It began as a multi-agent workflow in tmux, where Researcher, Architect, Dev and QA roles handed tasks to each other and checked each other’s work. It became a pixel-art virtual office: Claude Code, Codex and Gemini appear as little characters who work, walk around and hold meetings, and as the “manager” I assign tasks and can watch each agent’s terminal and conversation.

• Eight subsystems, including message relay across vendors, meetings, four-layer memory and a “dreaming” step that consolidates memory.
• About three months of work; the project is finished.

Demo (sample data)

A portfolio and brand site for Zongheng Brian He, a visual artist in Los Angeles, with five pages: Home, Portfolio, Club, Brand and Contact.

• Lately I’ve been building a tunnel-style portfolio gallery.
• Pushing code builds the site and publishes it to my own server automatically.

brianhe.studio

A workspace that keeps AI conversations, notes and references together. I designed the interface in Figma first, then built a pnpm monorepo with a React + TipTap editor and a separate back end.

Demo (sample data)

Tasks, plans, a focus timer and an AI assistant on one dashboard. I built it alone; the interface was generated with Figma Make.

Demo (sample data)

A web tool for improving AI prompts: one-click optimization with streaming output, multi-round deep optimization and scoring on the six COSTAR dimensions, with models from OpenRouter, OpenAI, Anthropic and Gemini.

GitHub

Shows how a campus uses electricity, forecasts demand, adjusts building equipment and demonstrates how the campus would respond to a demand-response request from the grid. It runs on simulated data, with energy and carbon reports.

An app that connects to Gmail and sorts out your schedule automatically. I built a first version of its website and wrote a business plan.

Joined a student service club of about 20 people. When I joined, its service activities were still being planned.

Quick prototypes made with AI site builders such as Lovable and v0: Cloud Chat (team chat with a Supabase back end), an internal admin dashboard, and early versions of Code Street.

Cloud Chat demo

A platform connecting developers with clients, with developer, client and admin roles, project posting, bidding, delivery and payment. My part was the first version of the front end in August 2025; teammates wrote most of the rest, and the project ran until January 2026.

Demo

A minimal community that posts one open question every midnight: no likes, comments or follower counts, just a “resonate” button. I also built a Next.js front end for a mental-health and self-exploration platform alongside it.

A diary for Apple platforms that takes periodic screenshots and records app usage, then sorts and analyzes them with on-device AI to help you look back on your day. I wrote the product spec (v1.2) and early Swift code; it stopped at that stage.

A one-week finance study program run with China CITIC Bank, the program’s 16th cohort: visits to financial institutions, sessions with industry mentors and hands-on practice, to learn about the industry and careers in it.

Reads the contents of PDFs with OCR (EasyOCR + PaddleOCR), renames them in bulk by rule, straightens rotated scans and backs up the originals, all from a web interface.

Demo · GitHub

Reads emotion second by second from the faces (DeepFace) and voices (wav2vec2) in a video, combines the two, and has a local language model (Ollama + DeepSeek) write music prompts to match.

README on GitHub

A home server that hosts this site, Brief, Studiolo and a few private services.

• Every release is kept separately and health-checked before traffic switches to it, so a bad release can be rolled back.
• Since August 2026 I’ve also run a 27-billion-parameter open-source language model myself, with llama.cpp on an NVIDIA DGX Spark.

Advanced to the Gold division of the USA Computing Olympiad (USACO) after about two months of practice.

I write and mix my own music in FL Studio. As of October 2026, the library on this site holds 40 pieces — 17 finished tracks, 16 demos and 7 samples — including some house.

• It’s also where Celadon and this site’s DJ player started.

Listen

Took part once in an in-person volunteer event supporting people with depression.

Painted with the art club for a semester, and my work was exhibited at school.

An early mental-health website I built. Both the code and the site are gone now.

Led a badminton club of more than 20 students.

Practice projects: a virtual pet in Python, Lovetern, a design for an AI recommendation system (Next.js + Prisma), and deploying an open-source ChatGPT bot for WeChat.

Virtual pet on GitHub

Generated images with Stable Diffusion, one of my first hands-on experiences with generative AI.

Played on my school’s badminton team from the spring of grade 9 to the fall of grade 10.