Jinhao Hu

Jinhao Hu, résumé portrait on a violet background
  • 1,471agent tasks, created Oct 4 to 24, done by Oct 28
  • 2,167real orders, Aug 1 to Oct 24
  • 86scheduled jobs enabled on Oct 28

I turn wild ideas into AI tools that make life and work easier.

AI engineeringagent workflows with review gates

MS in Information Management at the University of Washington, graduating June 2027.

I define what a system should do and AI coding agents build it. Work I flag for acceptance counts only after an independent AI review.

Open to
  • AI engineering
  • AI product management
  • AI operations and enablement
  • Business and systems analysis
  • Data analytics
Scroll
  • Claude Code
  • Codex
  • Dify
  • MCP
  • LLM-as-judge evaluation
  • Retrieval-augmented agents (keyword search)
  • Python (pandas, AI-assisted)
  • SQL
  • Git
  • Cloudflare Workers/D1
  • Multi-agent orchestration
  • Prompt workflows
  • Acceptance testing
  • Cursor
  • User flows
  • Axure RP prototyping
  • Playwright
  • MySQL
  • Google Analytics
  • Snowflake
  • Tableau
  • Dimensional modeling

01About

I study Information Management at the University of Washington, specializing in Business Intelligence and Data Science.

What I own is the rules around the work: what counts as done, who checks it, and what an agent may never do without me, like sending an email.

That is how a family building-materials store with 50 staff moved off handwritten sales slips, and how I run parallel worker agents under an orchestrator agent across my own projects.

Education

Master’s degreeSeattle, WA

University of Washington

Information School

MS in Information Management

Business Intelligence and Data Science

Sep 2025 to Jun 2027 (expected)

Bachelor’s degreeNewark, DE · Chengdu

University of Delaware

BS in Management Information Systems

Minor in Business Analytics

Joint program with Southwestern University of Finance and Economics (SWUFE-UD Institute, Chengdu)

Sep 2021 to Jun 2025

Parallel agents, not a chatbot

1,471

agent tasks completed across all my projects

Cards created Oct 4 to 24, 2026, done by Oct 28

321 of them were marked for acceptance, so each counted only after a separate AI reviewer checked it

I run many AI agents in parallel across my projects. Not every task needs the same bar: for the ones marked for acceptance, the reviewer checks the files, confirms the commit is pushed and runs the related tests.

Software a real store runs on

2,167

real orders

Aug 1 to Oct 24, 2026

97% entered by the AI assistant that reads the store’s Excel sheets (1,859 of 1,914 on Oct 16)

This is the Fuqing Inventory System, running the real stock and sales of my family’s building-materials store (50 staff), which used to work from handwritten sales slips.

Work that runs on a schedule

86

scheduled jobs enabled, 41 of them run an AI agent

Live cron list, Oct 28, 2026

97% of AI-job runs succeeded over 13 days (434 of 447)

Email triage and news digests run every day, and course briefs four times a week, on the open-source Hermes agent I run.

02Projects

01 · Personal AI product · Jun 2026 to present

AgentDeck

One desktop app where I talk to a lead AI agent, and it assigns work to parallel worker agents (AI coding sessions) for me.

See how it works AI Showcase
Problem
Running many AI coding sessions by hand means a window per agent and no record of who did what.
What I did
Defined the orchestrator-and-worker model. The hard part was usage limits: I designed failover so that when one account’s usage allowance runs out, another of three Claude accounts or Codex takes over (124 handovers in 21 days in October 2026, across my Mac and Windows PC).
  • 23AI sessions at once at peak on one Mac, usually about 11 to 12 on the Mac and 8 on Windows
  • 83%of 422 reviews done by a different model than the worker

When the AI got it wrong

  1. Oct 3, 2026

    What happened

    An auto-delivery bug typed a worker agent’s report into the lead agent’s input box and pressed Enter, sending my half-typed message.

    What I changed

    Text in the input box is sent only when a person presses Enter.

  2. Oct 8, 2026

    What happened

    The dispatcher dropped the model setting without a warning, so 6 agent sessions ran on Claude Haiku and reported finished work that was empty.

    What I changed

    Dispatch now stops with an error on any setting it does not recognize, and every agent run has to confirm which model actually ran (applied 16 times).

02 · Personal AI automation · Sep 2026 to present

Hermes Intelligence Dashboard

One private site for my email, AI news, videos, podcasts, reading and calendar, fed by scheduled AI jobs that run on the open-source Hermes agent.

See how it works AI Showcase
Problem
Daily email, AI news, videos, podcasts, reading and calendar lived in separate apps and inboxes.
What I did
Defined one private site that opens on my iPhone, iPad, Mac and Windows PC. The hard part is trust: three times a day AI jobs read every new email in two inboxes and turn it into action items, under a prompt that forbids sending mail.
  • 1,664emails turned into action items, Sep 16 to Oct 25
  • 97%of AI-job runs succeeded over 13 days

Hermes · one day, 00:00 to 24:00

A day of scheduled AI jobs, in one scroll.

Scroll and the clock runs from midnight to midnight. Each job lights up at its real time, and the phone shows what it sends. Feeds come from scheduled jobs of the open-source Hermes agent, which I use and integrate.

00:00PT

Real schedule · Illustrative output
00:00NightNothing is scheduled yet. The first job runs at 01:00.

03 · Freelance · family building-materials business · Aug 2026 to present

Fuqing Inventory System

A web inventory system for a store that used to run on handwritten slips.

Problem
A store with 50 staff ran on handwritten sales slips, and stock was barely recorded.
What I did
It has stock counts sales can check before an order, credit-sale balances, customer search by phone, and an AI assistant that turns the daily Excel sheets into orders.
  • 820customers on file
  • ~28orders a day over the 30 days to Oct 24

When the AI got it wrong

  1. Oct 2026

    What happened

    The old AI import step had booked discounts as payments.

    What I changed

    I proposed checking AI-imported orders row by row against the store’s own sheets. A full-database scan then listed 118 orders with the same problem, corrected on Oct 17 after I approved the fix.

See how it works AI Showcase

Demo copy with sample English data. The real store’s data never leaves China.

Fuqing demo screen. Home: today’s sales, what customers owe the store, and low stock.
Home: today’s sales, what customers owe the store, and low stock.
Fuqing demo screen. Stock: sheets on hand for each board, from the daily stock count.
Stock: sheets on hand for each board, from the daily stock count.
Fuqing demo screen. Customers: search, then see what each one has spent and still owes.
Customers: search, then see what each one has spent and still owes.

Demo copy with sample English data. The real store’s data never leaves China.

04 · Personal AI tool · Oct 2026 to present

AI Resume Tailor

Writes a one-page résumé for a job posting only from facts I have confirmed, then another company’s model reviews it blind.

Built on

10 real job postings across 5 role types

Tested on

20 job postings it had never seen

I defined the pipeline and its review rules; AI coding agents built it. A Claude Opus writer drafts from a confirmed fact bank, code checks facts, layout and wording, then GPT (via Codex) reviews it: a blind recruiter read of the page image and a sentence-by-sentence fact audit. Contact details are stripped before any model call. Each reviewer finding became a general rule, not a one-off fix.

  • 15 of 20 first drafts passed both blind reviews on postings it had never seen (Oct 25)
  • 2 independent reviews by another vendor’s model

05 · Personal AI workflow · Oct 2026 to present

ClassBrief

Turns the Canvas readings and deadlines for each class into a pre-class brief with links to its sources.

I defined the requirements and the writer-reviewer loop; AI coding agents built it. A writer drafts each brief from the Canvas readings and assignment requirements, and a reviewer checks course dates, week names and required-reading lists against Canvas. A draft that fails is rewritten before it is delivered to the course Drive folder.

  • 9 pre-class briefs for 2 graduate courses (Oct 5 to Oct 24)
  • 2 to 3 h of reading prep saved per class, my estimate

06 · Personal automation · Oct 2026 to present

Job Search Pipeline

Watches public job boards and screens new postings by role, experience, US location and work-authorization limits. It never applies for me.

  1. 1,961job boards
  2. ~95,000roles on them
  3. 232candidates in 3 days
Why 140 were screened out (Oct 24)
  • Experience42
  • Role direction27
  • Location24
  • Employment type22
  • Work authorization13
  • Posting closed12

I defined the screening rules; AI coding agents built it. It polls public ATS job boards on Greenhouse (802), Ashby (699) and Lever (460), prefers 0 to 1 years of experience, drops anything over 3, and records the reason whenever it screens a posting out.

  • 181 companies behind the candidates from its first 3 days

07 · Open source · Jun 2026 to Oct 2026

XDeck DIY

A free desktop app for Windows and Mac that puts five or six X (Twitter) feed columns in one window.

I set the requirements and Claude wrote the code; the first simple version took about 30 minutes. It needs neither X’s paid multi-column subscription nor the paid X API. I published it as a public GitHub repository and shared it on X.

  • 14 GitHub stars (Oct 25)
  • 82 release-file downloads across 4 releases (Oct 25)
Code on GitHub

08 · UW course project · Business Intelligence Systems · Spring 2026

Retail BI Data Warehouse

Two stores, one bonus pool: a star schema in Snowflake and a Tableau dashboard to decide which stores stay open and how to split the pool.

Built a Snowflake star schema with 7 dimension and 3 fact tables, wrote SQL views for 4 business questions, and a Tableau dashboard I presented with my group.

  • 104.97% of the 2013 target for Store 5
  • 76.16% for Store 8. Recommendation: keep both open

03Experience

May 2025 to Aug 2025 · Shanghai

Shanghai Yuepu Network Technology

AI Data Intern

Mapped the database fields for a co-built, AI-assisted Python report workflow that cut preparation time per report by an estimated 80% or more (a separate measure of the same workflow: once the data is pulled, it fills the team’s 44-page close-out template in under 5 minutes), and built a DingTalk alert that checks live ad projects every 30 minutes.

  • Python
  • Dify
  • Xiaohongshu Ads API
  • DingTalk
  • Axure RP

Jan 2024 to Feb 2024 · Shanghai

Shanghai Suanliyou Technology

AI Product Intern

Built Google Analytics dashboards on new, daily active and paying users, wrote Playwright scripts that check the company’s website for problems, and turned AI tutorials into screenshot guides for visitors.

  • Google Analytics
  • MongoDB
  • Playwright

Jul 2023 to Aug 2023 · Hangzhou

Tailong Commercial Bank

Data Analyst Intern

Wrote SQL on Alibaba Cloud MaxCompute over credit data on about 390,000 individual and 36,000 corporate customers, built non-performing-rate tables by industry across 25 branches and village banks, matched the branch totals to the bank-wide count at 99.87%, and co-presented a report on AI use at 6 peer banks.

  • SQL
  • MaxCompute
  • Excel

Also

  • Teaching Assistant, Database Management (MISY330)SWUFE-UD Institute · Feb 2025 to Apr 2025
  • Assistant to Student CounselorSWUFE-UD Institute · Sep 2023 to Jun 2024
  • Assistant, International Affairs Exchange OfficeSWUFE-UD Institute · Jul 2022 to Jun 2023

04Off the clock

Hover to pause, drag, or tap a photo

  • SeattleSkyline at golden hour

  • Stone Forest, KunmingLimestone pillars · 2023.06

  •  Aurora after midnight

  • Joffre Lakes, British ColumbiaArms open on a log

  • Whistler, British ColumbiaA suspension bridge at the summit · 2024.08

  • Tower Beach, VancouverSunset over the water · 2024.08

  • Stanley Park, VancouverBy the Brockton Point lighthouse · 2024.07

  • Victoria, British ColumbiaIn front of the Parliament Buildings · 2024.07

  • The Butchart Gardens, British ColumbiaUnder the rose arches · 2024.07

  • Shuanglang, Erhai LakeUnder an arch of flowers · 2023.06

  • Crater Lake, OregonAbove the blue lake

  • Florence, OregonSea stacks at dusk · 2025.11

  • Victoria Harbour, Hong KongIn front of the skyline · 2025.05

  • Victoria Harbour, Hong KongThe harbour lit up at night · 2025.05

  • NingxiaDesert camels

  • Zhangye Danxia, GansuRainbow mountains

  • Ngawa, SichuanSnow peaks above a frozen stream · 2023.12

  • Space Needle, SeattleDusk on the lawn · 2025.12

  • Downtown SeattleBetween the towers · 2025.10

  • University of Washington, SeattleCherry blossoms in spring · 2026.03

  • Snoqualmie Pass, WashingtonSnowboarding · 2026.02

  • Snoqualmie Pass, WashingtonA snowy valley in low cloud · 2026.02

  • New YorkManhattan at night

  • Times Square, New YorkThe lights at night · 2026.03

  • Liberty Island, New YorkLower Manhattan across the harbor · 2026.03

  • Columbia University, New YorkOn the sundial, Butler Library behind · 2026.03

  • Charles River, CambridgeBoston across the river · 2026.03

  • MIT, Cambridge, MassachusettsUnder the Great Dome · 2026.03

  • Niagara Falls, OntarioAt the falls · 2024.08

  • Toronto Islands, OntarioThe skyline and the CN Tower across the bay · 2024.08

  • The Bund, ShanghaiLujiazui lit up at night · 2024.01

  • Western Xia Tombs, NingxiaEarthen tombs under long clouds

  • NingxiaOn a sand dune

  • Lujiazui, ShanghaiUnder the Shanghai Tower at night

  • Erhai Lake, DaliOn a pier

  • Erhai Lake, DaliOn a rock in the lake · 2023.06

  • Erhai shore, DaliSunglasses and a bandana · 2023.06

  • Cangshan foothills, DaliFields under big clouds

  • Smith Rock, OregonA lone figure on a spire

  •  Under a full moon