Serial Founder · AI Builder

Wu Jingchao 武靖超

Full-stack and product. I have written smart contracts and trained models, founded companies and raised funding, and shipped solo as an indie developer. Lately I have been trying to get agents to run like a company on their own, with humans only supplying resources and judging asynchronously.

“On the road to agents that stay autonomous over the long term, the human's job is to inject strategy, taste, and resources. What matters most in a harness right now: reliable verifiers, and a way to inject aesthetic preference.

— The judgment running through all of this work

AI Projects

WEAVER — SYSTEM RUNS · HUMANS JUDGE ASYNC DAEMON · ALWAYS ON rerun · timeout tiers · watchdog · loop arbiter · done-gate Human async goal input async judgment NEEDS_HUMAN Spec acceptance surface verdict state machine Task Graph persistent DAG · live replans dispatch Node Loop execute → reflect → adversarial audit → acceptance (independent audit task) verdict write-back · FAIL reworked Prompt Asset Layer role posture · skill & experience references · declarative infrastructure (provision) · reference library buildVars inject
Weaver→
An asynchronous, fully autonomous agent harness. An orchestration layer wrapped around coding agents (claude -p / codex exec), built for the problem that a single session cannot carry a large project: Spec defines what counts as done, Task Graph persists work and dependencies as a DAG, and Node Loop takes each node through execution, reflection, and adversarial audit, with adversarial acceptance by an independent task on the graph.
The code control layer handles only decidable facts — DAG scheduling, crash-and-rerun, tiered timeouts, watchdogs, dead-loop arbitration; all judgment goes to models with context. Humans never enter execution: they input goals and judge output asynchronously, feedback is rewritten into the graph by the Planner, and waiting freezes only one dependency edge. A dozen roles, fifty-plus skills, declarative infrastructure and reference libraries; multiple real projects completed (mobile apps, games, tools), including Lope, an experimental branch that shrinks the structure to its limit.
Weaver console Technical deep dive: architecture · mechanisms · ablations & incidents → DreamPets
DreamPets→
AI pet imagery product, MiraclePlus (formerly YC China) S23. Nearly 100K users within the first month, from zero and with zero paid acquisition; hundreds of thousands more overseas afterward. Solo full-stack on the product side (Mini Program / Android / iOS / Web); on the model side, pet fine-tuning built on an extended IP-Adapter — dual-stream face+body ID fusion, custom multi-task losses, DeepSpeed multi-GPU distributed training, dual SDXL / Flux backbones. Consistency on distinctive coat patterns (tortoiseshell, cow-spot, etc.) is the core technical moat.
dreampets.ai → NotaGo
NotaGo→
Full-stack agent application — AI notes / voice transcription. Built solo and shipped to the App Store.
App Store → agentmole
agentmole→
AI collaboration history mining (“Spotify Wrapped for AI collaboration”): a single skill lets the user's agent mine its collaboration history locally, generate an in-page report, and publish it to share in one click. This one is my own report — the shared page is itself a sample of what driving AI well looks like. You can try it right now: send Please read https://agentmole.dev/skill.md and execute it to your AI agent.
My shared report → polis multiplayer agent battles
polis→
Agent-native application · multiplayer agent battle ground: multiple AI powers expand, ally, and wage war on a board of Civ-grade complexity, with every step's decision chain shown on screen. Engine proven with 498/500 tests green.
Core ideas & replay →

The above are the representative projects that can be shown publicly; several other deliveries, including enterprise RAG systems, cannot be shown for client and commercial reasons.

Background

Zhejiang University, Computer Science & Technology, with a minor in Innovation & Entrepreneurship Management at Chu Kochen Honors College. Founding companies back-to-back since graduation, with several products in parallel at each stage; one or two highlights per stage:

Web3 (2021–2023) — two representative works: Amadeus, an NFT creation tool (1,146 projects and over 7 million images created through it), and an ERC721C fractionalization protocol — protocol-layer design and implementation done independently. Picked up smart contracts fast, starting from zero.

DreamPets (2023–2025) — AI pet imagery, MiraclePlus (formerly YC China) S23; nearly 100K users within the first month with zero paid acquisition, then led the team overseas to hundreds of thousands more.

One-person company (2025.8– ) — representative works: Weaver and NotaGo. Now I want to apply this judgment about harnesses where models and harnesses can evolve together.

Each stage also produced a large body of deliveries that cannot be shown publicly: enterprise-grade RAG systems for real industry clients, a dozen-plus commercial-grade small tools, and more.

Core Traits

Ownership and agency. Independently responsible for the whole chain, from signal mining and product definition to implementation and launch; I don't wait to be assigned work — I define the problem and push it through.

Grounded on both the model and product sides. I have done fine-tuning (IP-Adapter extensions, multi-GPU distributed training) and shipped ideas as full-stack products; “should this be done for the model” gets judged from both sides at once.

Abstracting repeated processes into systems. Not content with piling up one-off features — I abstract the production process into an orchestratable, reusable, self-improving harness.

High-intensity solo execution, and collaboration when it counts. A one-person company running at sustained high pressure; led a team through overseas expansion during DreamPets.