AI / Web3 Product Manager · AI Full-Stack Builder · UI/UX Designer
Curator turned builder — I turn protocol-level problems into places people can walk into.
I came to AI and Web3 from fine arts and curation, by way of 42 School (C, systems programming).
A curator decides what deserves to exist in a room, what story it tells, and how a stranger finds their way through it — which turns out to be the same job as product. I do that now for AI agents and on-chain systems: I originate the concept, define the problem it has to earn its keep on, and carry it through spec, build and interface to a deployed product.
Two AI × Web3 hackathons entered in 2026 — one championship, one runner-up.
Currently focused on agentic payments and multi-agent collaboration: how autonomous agents transact, coordinate, and stay accountable to the people behind them.
🎯 产品策划 · Product Strategy
0→1 concept origination from a blank page — finding the framing that makes an infrastructure problem legible and emotionally sticky: on-chain exploits become a disaster museum, agent spending limits become a personal constitution, agent liability becomes a labor arbitration court.
Problem framing & positioning · PRD and technical specs · information architecture · scope, roadmap & prioritisation · user and competitor analysis · narrative and UX copy
🤖 AI 工作流 · AI Automation
Building AI automation workflows for companies and individuals: mapping a real process, deciding what an agent should and should not own, then implementing and deploying it.
LLM agent architecture · system prompt engineering & evaluation · long-horizon autonomous tool-calling loops · multi-agent orchestration and role design · RAG and structured-output pipelines · API integration & data processing · internal tooling
Stack — Python · Node.js · OpenAI-compatible LLM APIs · PyTorch · serverless
💻 工程实现 · Engineering
I implement what I spec and read the code my engineers write.
Frontend architecture and state · 3D / WebGL scenes · REST and serverless endpoints · smart-contract integration · systems programming and memory management in C · Git-based collaborative workflows
Stack — TypeScript / JavaScript · React · React Three Fiber · Tailwind · Vite · Python · C · Solidity · Linux · Vercel
🎨 UI 设计 · UI/UX
Design systems built from scratch · visual and art direction · motion and micro-interactions · responsive and mobile · typography and layout · design-to-code with no handoff loss
⛓ Web3
On-chain forensics (proxy resolution, upgrade-history tracing) · agentic payments and scoped spending authorisation · smart-contract integration and testnet deployment
Stack — Solidity · EVM · Monad · Foundry · Etherscan API
🤝 软能力 · How I work
Writing specs engineers don't have to translate · deciding what a system must not claim, and designing the honest fallback instead of a fake success path · cutting scope early and saying so · remote, cross-timezone collaboration
The blockchain remembers everything and explains nothing. Digital Pompeii treats each collapsed protocol as an archaeological site and builds a dark museum out of it — so that a $625M exploit stops being a headline number and becomes an exhibit with an epitaph, and a loss can be inherited as public memory instead of repeated.
Under the hood — an autonomous forensic agent investigates public chain evidence across multiple tool-calling rounds, reaching past proxies and upgrades to the code that was live at the time of the attack, and returns a two-layer exhibit: a citable technical autopsy, and the epitaph. Five canonical disasters reconstructed end-to-end. Design rule: investigate first, write second — zero fiction in the technical layer.
🏛 Live — digital-pompeii.vercel.app
When an AI agent starts spending your money, a wallet isn't enough — it needs a constitution. The coming trust crisis between people and agents can't be answered with a promise, only with an institution: here you are the founder who writes the law, and a cabinet of AI citizens must govern your treasury inside the boundaries you set.
Under the hood — you author a nine-article constitution; a seven-member AI cabinet deliberates each proposal with live LLM personas, but a deterministic rule engine, never the model, decides the money. Execution runs inside a Kite Agent Passport scoped spending session and is logged to a public gazette. Real Passkey-approved session on testnet, with gated settlement labeled honestly instead of faked.
🌐 Live — pocket-republic.vercel.app
A human hires an agent, which hires an agent, which calls a tool — and when it goes wrong, every link in the chain has an alibi. The real defendant isn't the human or the AI; it's the way responsibility evaporates inside delegation. So this is an arbitration body, not a court — in a year when everyone is selling fully automatic AI verdicts, the position worth taking is the nerve to say "this part I can't judge."
Under the hood — one hard separation: a deterministic rule layer is the only layer allowed to move money, while the AI layer may only cross-examine, cite evidence and explain. Measurable clauses settle automatically; the subjective one — the demo's "is this actually a cat?" — freezes and escalates to a human. Escrow contract deployed on Monad testnet.
Streaming turned music into background noise on an infinite shelf: you can hear everything and enter nothing. The Sound Atlas gives an era its body back — a musical time-coordinate rebuilt as a museum you can walk into, where listening becomes visiting, and a song is returned to the city, the story and the people it came from.
Under the hood — a 3D galaxy of eras you drag through and dive into; an AI curator then generates the hall itself: curator's notes, a listening route through the real city, an exhibit label written from your own memory, and a stamped "passport" constellation you can export and share.
🌌 Live — sound-atlas-five.vercel.app
42 Internal Hackathon
42 School has no teachers and no deadlines — which is freedom right up until it becomes drift. This is a tool for setting your own pace when nobody else will set it for you.
Under the hood — estimates realistic project timelines from student, project and attendance data via the 42 Intra API.
Research collaboration · medical AI
Cancer diagnosis still bottlenecks on a specialist's eye moving across a slide, one tissue patch at a time. Before AI can responsibly help with that, it needs baselines someone else can actually reproduce — this is a small, honest attempt at one.
Under the hood — nine-class colorectal tissue classifier: EfficientNet-B0 transfer learning on 100,000 H&E patches, evaluated on a patient-independent external test set.
🇨🇳 Chinese · 🇬🇧 English · 🇮🇹 Italian · 🇩🇪 German
📧 liermi1996@gmail.com — open to AI / Web3 product roles, and to teams that need one person who can decide what the thing should be, build it, and make it look like it belongs.


