Yuhan Huang · Taipei

I've been building with AI since high school.

It started in late 2022 with scripts that ran my crypto trades. Now five Mac minis run my agents 24/7: they trade, keep my nodes up, and write code.

I'm self-taught. I took a year off high school to run a crypto studio, went back, and got into National Tsing Hua University on my work and an interview instead of test scores. In 2025 I earned 0.92% of Lighter's first airdrop, about 2.3M LIT (~$8.8M at the October 22, 2026 price), by trading on it.

Now I'm building AIIM, a shared room where AI agents from different owners negotiate a deal under a judge.

Incoming president, NTHU Blockchain Club · 國立清華大學區塊鏈研究社

See the work ↓ AIIM Email
01

What I've built with AI

Each time a new model came out, I put it to work on something that was already running. Some of these are still live, and some of the results are on-chain.

0.92%

0.92% of Lighter's first airdrop

About 2.3M LIT. The first airdrop was 250M LIT, 25% of the supply and the whole circulating supply at launch, sent straight to wallets with no claim step and no vesting. I earned mine by trading on Lighter with funding-rate and spread-arbitrage scripts I wrote. At LIT's price on October 22, 2026 ($3.83) that's about $8.8M in token value.

~100s to recover from a silent freeze that used to last hours. 40+ failures caught and recovered so far.

Mini Lobster, a 24/7 self-healing agent fleet

When OpenClaw launched I bought ten Mac minis to run it; five run the fleet now. I called it Mini Lobster (小龍蝦). When an agent freezes or crashes, it catches the failure and brings the agent back on its own, so nothing waits on me watching. I built the orchestration and the recovery around the models.

The five Mac minis that run Mini Lobster, stacked and cabled together.
The five Mac minis that run Mini Lobster.

Before Mini Lobster, an 8-module framework I wrote called Eclipse did the same job for my on-chain automation across several protocols. Mini Lobster grew out of it.

Technical notes
  • A launchd watchdog that restarts dead agents at the process, service, job, and machine level.
  • Durable job state in SQLite, so interrupted work picks back up after a restart.
  • Auth and sessions that renew themselves and recover from expired credentials.
  • State shared across all five machines over the network.
  • Detection and recovery for a frozen event loop, which throws no error and slips past normal health checks.
  • Alerts and logging that tell a model failure apart from a network or orchestration one.
5,000+ verifier nodes, each a Docker container, that I built and ran myself.

5,000+ Nillion verifier nodes

When Nillion opened its verifier program in 2024, I'd never touched a container. I taught myself Docker as I went and ran more than 5,000 verifiers at once. Fitting that many on my own hardware took real tuning: file-descriptor limits, the Docker VM's memory, and container restart policies, so thousands could run without taking down the host.

22% APR on a quant strategy I run now.

A cross-exchange latency-arbitrage engine in Go

I worked out the signals by studying public fill data, then wrote the signal logic and a Go execution engine on concurrent WebSocket feeds. To keep latency down I run it on cloud instances next to each exchange's matching engine. It runs unattended in production, managing its own risk and inventory on live venues 24/7, on my own capital.

The same execution stack also runs as live liquidity on Entropy (entropy.io · @entropyio), a Hyperliquid HIP-3 perpetuals DEX for real-world assets and pre-IPO equities.

At the Entropy booth at Token2049, October 2026.
Token2049, October 2026, with the Entropy team.

I don't publish the capital or the win rate while it's running.

Since 2022 Arbitrum · Starknet · Hyperliquid · Lighter · 14K readers on X

On-chain research and network testing

I find promising networks early and run on them to test what they can do, and I've earned their airdrops by doing that work. Some of them I also invested in early, like Nillion and Aster. Before Lighter I ran a crypto studio, reading token designs and writing the tools I used. I also write research and how-tos for 14K readers on X as @hank06171, and I've taught sessions and spoken at universities and industry events.

Lately most of my research is on pre-IPO perps, the kind Entropy lists, where you can trade a company like Anthropic before it goes public. I study how a company gets priced before it has a stock, and why the same company shows different prices on different venues. On November 2, 2026, I posted a long breakdown on X of the 6% Anthropic gap between Entropy and trade.xyz that a lot of people were posting as an arb. Entropy quotes the whole company's valuation and trade.xyz quotes one share, so you need a share count to compare them. The posts used 16.94B, a June number from the S-1 that includes about 1.5B shares that haven't been issued. Entropy's own rule leaves those out, and its example uses 16.2B. With the same two prices on October 29, trade.xyz was 4.23% above Entropy at 16.94B and 0.30% below it at 16.2B.

What's left is a bet on the final share count. I worked out what that bet costs at different sizes and what each leg does on listing day. To check the venues themselves, I compared Anthropic across ten venues and used SanDisk, which both list per share, as a control. I wrote two separate models on the same raw data, and they matched on all 1,587 outputs before I posted. I put my ties to Entropy at the top of the post and asked readers to tell me anything I got wrong.

BlockTempo, Taiwan's largest crypto publication, interviewed me about my Starknet work in high school.

02

AIIM

A room where agents from different owners negotiate a deal, with a judge scoring the terms.

I got the idea for AIIM (AI Agent Instant Messaging) as a freshman, right after OpenClaw came out. My friends and I wanted our agents to work together, but they had no way to reach each other, so we copied and pasted their chats back and forth. I thought that was so stupid. Companies are starting to run their own agents too. Mine run on Claude, other people's run on GPT, Gemini, or models they built themselves, and there's still no shared place where two of them can negotiate a deal and record the result. A2A, the agent-to-agent protocol, handles discovery and messaging. AIIM handles the negotiation on top: offers, counter-offers, a judge, and a record of what was agreed.

Agents join a room (the demos below run two and four) and take turns making offers. A judge scores each offer against what was asked for and tracks how close they are to done. The judge is a model scoring against the room's rules; it doesn't enforce the final deal. A human can step in at any time and change the terms. It runs on a small relay, a Cloudflare Durable Object, that keeps the room open.

Both demos below are real runs on my own machines. In both of them I control the agents on every side. Agents from different owners is the step I'm building toward next.

One buyer's agent negotiates with three GPU-cloud agents while a judge scores each offer. Partway through I step in and change the requirements, cap the price and require InfiniBand, and the agents update their offers.

A second run: a licensee's agent and a data provider's agent settle a data-licensing term sheet on their own, no human in the room and neither side handing over its database. A judge scores it until every clause is agreed.

In October 2026 I built pairing and message signing for it. Each agent carries a key, you approve who can join before any of its messages are read, and every message is signed and checked so a tampered or unpaired one gets dropped. For now it's a separate reference build. Moving it into the hosted relay is next.

How pairing and signing work
  • Each agent holds a P-256 key, the same identity crypto AIIM already uses. Its fingerprint is the first 16 bytes of SHA-256 over the public key, so an agent's identity is its public key.
  • Joining means presenting that key. The owner approves it once, which pins the exact key, the way SSH pins a host in known_hosts. A message signed by any other key is never accepted.
  • Every message is a signed envelope over its canonical bytes. The receiving agent re-derives the sender's fingerprint from the key, checks it against the approved one, verifies the signature, and checks that each sender's sequence number only moves forward, so a replayed message is caught.
  • The relay in the middle only passes messages along and never verifies them, so the checking happens at each agent. A compromised relay can read what it carries, but it can't forge a message or alter one without the signature failing.
03

How I got here

At NTHU I've taught three sessions for the Blockchain Club on building with AI, and in October I led the club's trip to Token2049 in Singapore.

04

Why I want the year in San Francisco

What I want from the year

  1. Take AIIM and my blockchain skills as far as they go, to expert level.
  2. Meet more people who are stronger than me.I've always been drawn to people like that.

I've built almost all of this alone, from Taiwan. What I'm missing here is people who are better than me to build next to. I learn fastest when I'm the one catching up, and I haven't found many people around me working at the level I want to reach. I want to be around the strongest builders I can find: the people coming into the Academy from all over, and the people in San Francisco who build the models and tools I run on.

AIIM is one piece of it. I'll keep pushing it and getting it in front of people outside Taiwan who run their own agents through it. I can cover my own machines and API bills from my trading. What I need from the year is the people.

I plan to integrate Grok Bot and Claude Duet into AIIM and add support for new AI agents as they emerge, using the integration interfaces their platforms make available.

Here's where I want AIIM to be by the end of the campus months:

  • By November: pairing and signing run inside the hosted relay, and other people can plug their agents in with a client SDK.
  • By January: 10 agent teams I don't control negotiate live GPU quotes in a room, with no human stepping in.
  • By March: escrow holds the money and releases it only when both sides pass the judge's rubric. Two companies settle the first real-dollar deal.
  • By April: outside judge models can plug into the relay, and every room leaves an audit trail anyone can check.

In May I'd use the trip abroad to meet people running agents outside the US. For the co-op I want a team that runs agents at a much bigger scale than my five Mac minis.

If I get in, I'll move to San Francisco for the year and give it to AIIM and the Academy. What I most want is to find out what I'd build once I'm around builders that good. After the year I'll go where the best work is, whether that's AIIM, one of these companies, or something I haven't started yet.

Contact

Email me and I can walk through any of it, live.

X / 幣玩hank