Skip to selected work
Harvey Zheng

Harvey Zheng

Product designer and frontend engineer. Experience building AI/ML products across AWS Bedrock and SageMaker.

Carnegie Mellon — HCI and Statistics/ML.

Selected work

5 projects, 2022–2026

Curated selection, not chronological.

01

Plover — readiness & load intelligence for runners

Solo designer & engineer · Private beta, self-tested

  • Backtests run the same product code the app ships.
  • Feed index: 1.1 MB → 28 KB. Cold start: 17 s → 3.2 s.
Plover run journal: a run card with distance, pace and heart rate beside a performance signal, and a summary of the checks run on recent training.

A running-readiness product I took from a single Jupyter notebook to a deployed app with a marketing site in about twelve days. It reads the data your watch already collects and tells you the two or three things that changed this week — and is deliberately honest that it does not predict injuries.

02

AWS Bedrock IDE: Model picker

UX Designer · Shipped in Bedrock IDE

  • Provider neutrality ruled out a curated "recommended" list; Spotlight plus recently-used instead.
  • Model metadata moved inline, with type-ahead and arrow-key navigation.

Redesigned the model picker for Amazon Bedrock IDE, cutting model selection from 5 clicks and a scroll down to 2 clicks across playground, agents, prompts, prompt flows, knowledge bases, and model evaluations.

Password required

03

A chat UI primitive, validated

Frontend Engineer (research + prototyping) · Research artifact; engineers built the package

  • Live theme switching across shadcn, Material and Ant Design — judged as real tooling.
  • Engineers built the package; parts now run in the Bedrock playground.

Validated an internal chat UI primitive by building a Wizard of Oz docs site for it before it existed.

Password required

04

SageMaker Canvas: Project vision

UX Designer (Intern) · 12-week intern project, shipped

  • Spent the final weeks on image sorting, not wider flow coverage.
  • Four engineers and a Sr. PM; shipped the quarter after re:Invent.

Designed a no-code CV model tuner for SageMaker Canvas, simplifying the complex process of fine-tuning image classification models into a platform accessible to non-technical business analysts.

Password requiredView AWS blog (opens in a new tab)

05

Anagnori — a self-guided exposure therapy prototype

Solo designer & engineer · Self-tested iOS prototype, not clinically validated

  • Built a clinically faithful v1, caught myself avoiding it, rebuilt around experiments.
  • Debrief redesigned so the pattern and prediction stay visible while you grade.
Anagnori: a pattern card with an avoidance name, values tags and an intensity slider, beside an experiment card tracking trigger, prediction, observed intensity and takeaway.

A self-guided iOS app I built around a framework that helped me in therapy. It started as a Google doc of experiments I was running on myself and became an app once keeping that doc by hand got old. The evidence behind it is my own use — it has not been trialled or clinically evaluated.

From the lab

Browse the lab

Solo prototypes I build with AI tools — 11 experiments so far, newest first.

Web Effects Pack

Fifteen film and print effects — burns, filmstock grades, crumpled paper, neon, a street light — running as real CSS/SVG/WebGL over live DOM instead of baked into a render.

Demo ↗

Plover

A running-readiness app I took from a Jupyter notebook to a deployed private beta in about twelve days, leaning hard on AI agents. Built multi-user; so far I'm the only one on it.

Demo ↗

unit-of-trust

A scrollytelling essay arguing that agent decision legibility should be a first-class layer of the agent stack, not an afterthought.

Demo ↗