Hanyu Liu — AI product & systems design

Designing AI systems for real‑world decisions.

From ambiguous inputs to structured reasoning, deterministic execution, and human-verified outcomes.

Hanyu Liu — AI product and systems design. Architecture and spatial research first, then manufacturing operations, now AI systems built and validated on real work, by one person working with AI coding agents under direction.

Featured caseSelected systemsAbout

Selected systems

Four more systems, each built around a different kind of reasoning
Prototype, software-tested Separation workbench: an illustration on black fabric split into seven plates, with the agent's plan, quality checks and a halftone zoom Hover or tap: plates, compare, halftone zoom

Screen Print Separation

An agent proposes the plan; a person edits any parameter; the pipeline writes the production files.
4 separation modes, 7 platesspot, gradient, simulated process and indexed
In use
Career OS dashboard A job posting scored on four criteria with a decision row Resume Studio with the fact-check verdict shown on a tailored draft
Hover or tap: discover, check, track

Career OS

The same method turned on my own job search: scheduled discovery, scoring, a fact gate before any resume leaves.
5 stages, 1 gateno resume leaves with a fact it did not have
Interactive prototype 1B shoulder, point to point 46.0 3A chest, half 54.0 (×2 = 108.0) 9C body length from HSP 68.0 12K sleeve length from HSP 58.0 net sleeve from SP= 12K − 1B × 0.5 = 35.0 HSP SP TECHNICAL FLAT · FRONT · cm · synthetic size Mnot to scale Hover or tap: resolve the sleeve

SizeSpec, geometry reasoning

Where does a measurement start? Candidates are derived from the drawing and scored; the choice is explained.
3 candidates, 1 chosenevery number carries its derivation
Delivered, confidential Hover or tap: re-weight

Geospatial Intelligence

Census, mobility and point-of-interest records turned into a scoring framework for site selection and growth.
300K+ records, 10+ indicatorssite selection for 20+ growth companies

Design and experiments

Where the design background shows
Live demo A 3D teamwear kit rendered in the browser from a measured pattern Move or drag to rotate

Teamwear Digital Twin

A measured pattern simulated in CLO3D, exported to glTF, rendered in the browser.
In production Hover or tap: time flows

Order Node Reminder

Production nodes that warn before they are late and escalate when they are.
Shipped A compostable straw, from research to backers.

POCOO Green Life

A consumer brand taken from user research to a funded launch.

How I think

Three habits that recur in every project

Design for uncertainty.

Decide first which judgments may be probabilistic and which must be exact. In the quoting system the model reads and proposes; the arithmetic never leaves the rule engine.

Build for reliability.

Nothing the model says becomes a business fact until it is checked against evidence. Unsettled items are held for a person; every reviewer finding becomes a rule and a regression case.

Validate in the real world.

Measure on real orders, with the error defined narrowly enough to mean something: system computation error, net of human-set parameters and business terms, on quotes a person reviewed line by line.