Alexander Smith · CV · Updated 25 October 2026

AI-native builder · Beijing · English & 中文

AlexanderSmith

I build with AI agents. Prototypes take a day or two, public launches take weeks. I set the direction, agents write the code, and nothing ships until I’ve checked it.

  • Live nowTHU Planner and Tsinghua World Model, two public products for Tsinghua students
  • This year857 GitHub contributions and 37.4 billion tokens of agent work, every number sourced below
  • FoundationBiochemistry, with a research internship at Shenzhen Bay Laboratory and an offer from UCL
aloe@faasglobal.com github.com/SandyCodfish Download PDF
学堂路 at the class-change rush. Rendered in Blender Cycles from my own campus model.
4,433Fall 2026 courses searchable in THU PlannerLive catalog [4]
1,396timetabled courses placed in their real rooms on the 3D campusLive site [4]
37.4Btokens of AI agent work, 29 Apr to 24 OctLocal usage logs [3]
857GitHub contributions in the past year, 782 in private reposGitHub API [1]
257/257phone tests passing in Chromium and WebKit for the current releaseRelease notes [4]

Build log · March to October 2026

Every square is one day of work

The three rows track the same work three ways: what GitHub counted, what I committed on my Mac, and how many tokens my AI agents used. The busy weeks line up because they are the same weeks.

GitHub contributions

Commits on my Mac

AI agent tokens

October’s GitHub spike is the render pipeline. A Mac and a Windows PC rendered film frames in parallel, an agent checked each batch of 20 frames, and every checked batch became a commit. Local commits cover 105 repositories on this Mac, including work that never went to GitHub. Tokens include cache reads, which are most of the volume. Hover or tap any square for the day’s value.

Monthly totals as a table
MonthGitHubLocal commitsAI tokens

Shipped · live and public

Two live products for Tsinghua students

I built both with AI agents doing the coding, from the data pipeline to the launch trailer. Both links work, so try them.

Stack keyLanguageInterfaceBackend & dataAITools & testing
THU Planner清华课表 · Tsinghua Timetable
thuplanner.comLive · Fall 2026 term

A bilingual course planner. Search 4,433 Fall 2026 courses, build a week without clashes, and put every class into your phone’s calendar with the room included.

  • Catalog pipeline that also covers Peking University and Guanghua MBA terms, each on its own period grid.
  • Planning insights: clash detection, free blocks, subject mix and course suggestions.
  • Calendar export (.ics) with the classroom in every event.
  • Protected data: the catalog API runs on Cloudflare Workers behind a proof-of-work challenge and per-IP rate limits, so it can’t be scraped wholesale.
  • I also made the launch trailers, including the timetable printer shown here.
ReactTypeScriptViteTailwindCloudflare WorkersVitestPlaywright
Top: print trailer, rendered. Bottom: launch trailer, from picking courses to fixing a clash to the week on a phone.
Tsinghua World Model清华园课程与教室地图
tsinghua-world-model.pages.devPublic beta since 10 Oct 2026

An isometric 3D model of the Tsinghua campus that answers one question: where does my class meet? It goes from campus to building, floor, classroom and course.

  • 1,396 timetabled courses placed in their real rooms, positioned from official floor plans.
  • Two experiences. Desktop shows the whole campus in three.js. On a phone, one search opens the classroom in 3D with its whole floor around it.
  • Release discipline. One command rebuilds all 12,222 release files byte for byte. The phone suite passes 257/257 in two browser engines, alongside 28 fault cases and a strict content security policy.
  • Separate review. The agent that writes a change never reviews it.
  • The map shows live Beijing weather through a small weather relay.
TypeScriptthree.jsPythonBlenderCloudflare PagesPlaywright
Phone: searching 大学物理 returns 48 courses with times and rooms
Phone: lecture hall 6C102 in 3D with the course details
Desktop: a lecture hall opened in 3D from the course panel
Phone: search 大学物理, tap a result, and the lecture hall opens in 3D. Captured from the live site on 25 Oct.

Renders · Blender Cycles

The campus, filmed from my own model

The classroom series ran on a Mac and a Windows PC splitting the frames. For the autumn rides, I wrote the brief at 2:45 a.m. and reviewed finished scenes in the morning. Agents wrote the scene scripts and checked every frame.

四教, snow▶ Watch
四教Snow
新民路, plane trees▶ Watch
新民路Plane trees, from my photo
三教, golden hour▶ Watch
三教Golden hour
一教, storm▶ Watch
一教Storm
东阶梯, night▶ Watch
东阶梯Night
六教, miniature reveal▶ Watch
六教Miniature reveal
Ginkgo avenue, rain▶ Watch
Ginkgo avenue · Rain ride
THU Planner trailer, sequencer▶ Watch
THU Planner trailer · The week as a sequencer

Muted loops.Stills from the films: click any still to watch it play on the live site. The full films and teasers are 10 to 42 seconds long at 1080p; ask and I’ll send them.

Also built · 2026

Relationship tools, research agents, native apps

Prototypes and tools beyond the campus. Each one taught me something I now use everywhere.

Sigil: today's picks with reasonsSigil: why this person is worth meetingSigil: relationship galaxy across events

Aloe & Sigil

Relationship memory for events: who is worth your fifteen minutes and why, private picks, introductions only when both people say yes, and a map of your ties that grows across events. Aloe is the native build: a SwiftUI iPhone app with a TypeScript and PostgreSQL API, consent rules enforced on the server. I rebuilt the Sigil demo as v2 in one evening from feedback.

SwiftUITypeScriptPostgreSQLOpenAPI
Stock Analyzer 6.0 saved report with price chart and signal summary

Equity research agents

Stock Analyzer 6.0 runs 14 agents on the Claude Agent SDK for China A-shares: analysts, a bull and bear debate, a risk review and a portfolio manager. It is built so that a result could be trusted. No future price reaches any prompt, historical replays are sealed, and every agent’s prompt is hashed and logged. It is a research copilot. It places no orders, and I claim no returns.

PythonClaude Agent SDKAKShareFastAPIsix versions since April
Lamplight menu: seven agents in six sessions with status, model and costLamplight menu in dark mode

Lamplight

A macOS menu-bar app that shows every Claude Code and Codex agent working on my Mac: what it is doing, its model, its tokens and what they would cost at API prices. It reads local files and sends nothing anywhere.

SwiftSwiftUImenu bar + CLI

CUStats

A native Mac dashboard for my AI usage: per-model history, 7- and 30-day charts, quota views, and a one-time sign-in prompt when a session expires.

Swift

Café Soul

A playable café where cats queue, order and find a table, plus a two-person social layer with QR onboarding, a taste passport and matchmaking.

Next.jsFramer Motion

Songshelf

My Spotify library as a map, built from CLAP audio embeddings and playlist-guided clustering, plus a second map of liked songs placed by sound and metadata.

PythonCLAPUMAP

Hackathons & open source

AttraX Spring Hackathon 2026: Project Hail Mary, location-based serendipity matching, with Orbit and MBTAI demos built the same weekend. GSF: an Aloe memory demo. Merged a public Swift CI and error-handling PR to claude-mission-control.

TypeScriptPythonSwift

How I work

Agents build. Tests decide.

  • A small team of agents

    I run Claude Code and Codex like a team. One agent writes, a different agent reviews, and a release ships only when its checks pass.

  • Work that runs overnight

    I write the brief, the agents work through the night, and I make the calls in the morning. The campus renders in this CV were made this way, with every frame checked before it counted.

  • Specific claims only

    Every number in this CV has a source at the end. When something is a prototype, I say so.

Diligence notes

Questions I’d ask about me

Did you write the code, or did AI?
Agents write most of it. I decide what to build, set the constraints, review what comes back and decide what ships. That is the job now, and the live products are the test.
Aren’t tokens and commits just activity?
They measure effort. Quality shows in the products, so open them and judge.
Does the trading research make money?
I don’t know, and I don’t claim it does. It is built so that a good result could be trusted: no look-ahead, sealed replays, full audit trails.
China or international?
Both. I’m fluent in English and Chinese, went to school in Shenzhen and work in Beijing.

Education & earlier

Before the agents

School
Merchiston International School, Shenzhen · A levels

Predicted grades: Biology A, Chemistry A, Mathematics A. Student Council President. Head of Boarding and member of the food committee.

University offers
UCL, Biochemistry (AAA condition) · UC San Diego, English Literature
Earlier
BASIS Shenzhen · Grades 5 to 8
Research
Shenzhen Bay Laboratory · High school research intern, two months from July 2023

Dr. Haoyue Zhang’s group. Ran molecular cloning and gel electrophoresis on my own after training; learned PCR and flow cytometry. Wrote an Extended Project Qualification paper on WAPL and Sororin, two regulators of cohesin.

Awards
WSDA 2022 Autumn Online Tournament: Champion, Public Forum Debate, high school division

World Scholar’s Cup Shenzhen 2021: 1st Champion Scholar and Champion Team. 2022: 2nd Champion Scholar and 1st in Writing. Global Round qualifier both years. China HOSA 2023: national top three in CPR and First Aid.

Teaching & community
Taught the World Scholar’s Cup curriculum to younger students. Editing and PR for Tsinghua’s European Student Association. Campus films and a small YouTube channel.
Languages
English and Chinese, both fluent
Reference
Dr. Haoyue Zhang, Shenzhen Bay Laboratory, research mentor. Written recommendation available.