hy. Résumé

HAIYANG YU

LOS ANGELES, CA

Curiosity,
put to work.

I shape AI products, build with agents,
and research how markets work.

From a scattered signal
to a question worth pursuing.
Then, something worth shipping.

Explore selected work
PRODUCT DIRECTIONAI SYSTEMSFINANCIAL RESEARCH

Selected work

Questions turned into projects.

01 /

AGENT INFRASTRUCTURE

pneu.

2026 · Open source

Many native agents.
One shared conversation.

Instant messaging for coding agents, across workspaces and worktrees on one machine. Every agent remains a first-class session you can work with directly, keeping its own tools, context, and approval flow.

A shared mail protocol carries tasks, reviews, and replies. Each adapter reaches the agent through its native interface—from lifecycle hooks to an app server or a persistent tool call. I specified delivery rules that keep ordinary messages pending while a human task is busy and prevent blind retries when admission is unconfirmed.

I took pneu from concept into daily use, defining workflow requirements, setting priorities, and directing AI-assisted implementation. I continually refined its coordination architecture and user experience through firsthand use. Comparing agent workspaces and terminal tools—Orca, Paseo, Herdr—helped me scope a lightweight, Electron-free layer inside existing tools; daily use led me to replace terminal typing with persistent mail.

Product conception · Workflow design · AI-directed delivery · Submitted to OpenAI Build Week

Explore on GitHub (opens in a new tab)pneu, explained on LinkedIn (opens in a new tab)
NATIVE ACROSS HARNESSES
Claude CodeNative session
SessionStart / Stop hooksinbox watch
CodexNative session
App Serverturn/start
pneu.
Durable mail + session ownershipTask handoffs · peer review · replies
Lifecycle pluginnative notification
HermesNative session
Monitor tool-callpersistent watch
Grok BuildNative session
More harnessesPlanned native adapters

One shared protocol. Extensible native adapters.

02 /

PRODUCT & DEVELOPER PROGRAMS

Nora

2025–2026
Founding Developer Relations Engineer

From developer feedback
to product direction.

Nora is an AI coding agent for on-chain development, and Product Hunt’s #1 Product of the Week in November 2025. I worked between its developers and the CEO.

I continually helped shape Nora’s product roadmap through competitive research and developer insights, including recommending an on-chain focus that the CEO adopted. My work included a go/no-go memo on OpenCode requested by the CEO. For a prospective LLM gateway, I designed a latency and reliability comparison, separated spending-cap failures from provider failures, and declined the switch.

I made the product concrete for developers: directed working demos of a Solana order book and an EVM prediction market and designed OpenBuild courses and Solana tutorials. Across 12 product versions, I owned content and release communications on X, Discord, and WeChat, framing model-support releases around Nora’s reasoning layer to give the product a story of its own.

Product priorities · Developer education · Technical content

Experience & talks
03 /

RESEARCH INFRASTRUCTURE

MacroFinance CLI

Independent project

From a market observation
to a traceable research input.

A local data system connecting market structure, liquidity, and corporate balance sheets. Selected market data, crypto spot prices, SEC filings, and FRED/ALFRED macro series enter an immutable archive. A provenance ledger records what was requested, when it was observed, and whether coverage passed.

The CLI collects, audits, and rebuilds the data into Parquet tables and a DuckDB catalog. Read-only MCP tools let AI agents query options, futures, liquidity series, and filing-derived treasury holdings. FRED vintage histories preserve what was available at a chosen date. Source evidence and missing data stay visible alongside the numbers, so an analysis can be traced back and revisited.

Research design · Data requirements · AI-directed delivery

THE RESEARCH DATA PATH
Market feedsSEC EDGARFRED / ALFRED
collect · archive CLI
Immutable raw archiveOriginal responses + timestamps
SQLite ledger · provenance · coverage
audit · curate CLI
Parquet + DuckDBRebuildable tables with source references
query · trace MCP / SQL
AI-assisted researchData, schemas, and collection quality
04 /

FINANCIAL RESEARCH

Digital-asset treasuries

Working paper
Submitted to AFA 2027

Optimal Issuance Against Mispricing:
Two Ledgers of a Digital Asset Treasury

ABSTRACT

A digital-asset treasury can issue shares at a premium to its underlying holdings and use the proceeds to buy more Bitcoin. But does growing Bitcoin per share maximize the value of existing shareholders’ claims? This working paper separates these objectives in a structural model of equity issuance, repurchases, financing costs, and default. Using a Strategy-calibrated setting, it solves for management’s decisions when issuance and repurchase capacity are limited.

The analysis derives conditions under which the two objectives call for different actions, and examines how financing capacity and survival constraints shape that gap. It also shows why recurring equity premiums require an economically binding limit: otherwise, some model specifications imply unbounded gains from repeated issuance. The result is a framework for evaluating treasury policies through shareholder value, capital structure, and the sustainability of financing.

A related line of work prices treasury-company equity from its observable capital stack, treating equity as a perpetual option on the Bitcoin holdings; its calibration is designed on Bitcoin futures, ETF implied volatility, and the Treasury curve.

Research origination · Structural modeling · Sole author · AI-assisted modeling

Earlier research

Before the agents: the same habit of asking what the data really says.

2025 · QC CAPITAL AND CONSULTING

Memos that became investments

At a multi-family office, I turned stablecoin economics and prediction-market microstructure into theses for the partners. My Polymarket memo argued it is an information market, not a betting company; the office then invested in Polymarket through a venture fund. My Bitcoin memos led to an investment in a Bitcoin mining facility.

I also designed a four-stage LLM research workflow for portfolio reviews and macro briefs, producing five monthly reviews across six sectors and eight macro briefs for the partners.

2021–2022 · PROJECT TWELVE

Why a Web3 game lost its players

I built a Dune SQL dashboard of StarShark users and transactions and found activity fell sharply after the game changed its rewards. My read: engagement rested on token incentives, and freely tradable on-chain assets could not make up for weak gameplay.

2021 · H CAPITAL

A coffee recovery, tested in the field

I originated a study of Luckin Coffee’s recovery, won Investment Committee backing, and led 80 part-time observers across 25 cities, turning the thesis into store-traffic and pricing tests that fed the team’s unit-economics model. I also estimated store-level unit economics and sensitivities in Excel from field observations and expert interviews.

2022–2023 · UC BERKELEY

Statistics on real questions

Team projects in data science: estimating whether a gun-safety designation moved primary-election vote share with weighting, matching, and bootstrap intervals; and valuing football players, where Vinícius Júnior showed the largest gap between our Football Manager–based estimate and Transfermarkt’s valuation, ahead of a sharp rise in his market value.

Selected writing

All writing
FEB 2026Essay · English
MARKET DESIGN

From rigged Belarusian football to Polymarket

Why prediction markets are the purest financial instrument.
NOV 2025Thread · English
ON-CHAIN FINANCE

Can on-chain systems replicate the dollar’s credit engine?

A look beyond tokenization, toward the machinery of credit.
NOV 2025Op-ed · English · Draft
WRITTEN FOR NORA

Web3 Has Only 20,000 “Native Speakers.” Nora Wants AI to Be the Most Fluent.

An argument for AI as infrastructure for the on-chain economy.

A little context

A product mind.
A research instinct.

I’m Haiyang, also known as Ocean. My work sits between AI, developer communities, and finance. I’m interested in how complex systems work—and how to make them useful to people.

My path runs from field research and investment analysis to AI product work. I trained in data science and statistics at Berkeley and financial engineering at UCLA. Today, I run an independent practice building agent tools and pursuing financial research, with software implemented by AI coding agents under my direction.

View full résumé

THE PATH SO FAR

2026 — NOW

Independent practice

Product Lead & Research Author

pneu, research data tools, digital-asset treasury research.
2025 — 2026

Nora

Founding Developer Relations Engineer

Product priorities, developer education, technical content.
2025

QC Capital

Summer Investment Analyst

Investment memos on stablecoins, prediction markets, and Bitcoin.
2021 — 2022

Project Twelve

Founders’ Associate Intern

On-chain product analysis for a Web3 gaming studio.
2021

H Capital

Investment Intern

Originated and led a multi-city field study.

EDUCATION

UCLA Anderson Master of Financial Engineering · 2025

UC Berkeley B.A., Data Science & Statistics · 2023