Work · Zixuan

In this selection

Independent product

Roommate matching system for shared housing

Structured intake, staged matching, masked contact handling, and agent-side workflow support for student-focused shared housing.

The demand appeared repeatedly in Boston's student rental market, especially around shared housing and roommate search. People were not only looking for apartments. They were trying to align lifestyle, budget, move-in timing, and compatibility before even reaching the application stage.

What existed in practice was noisy and unstable. Most roommate discovery happened through Xiaohongshu and similar social channels, where posts were easy to miss, difficult to compare, and quickly pulled into private chat before preferences could be organized. Discovery was only part of the problem. The larger issue was coordination.

The product response was a WeChat mini program designed and built independently from definition through front-end implementation. Instead of leaving the process open-ended, the system turns roommate search into a structured intake and staged matching flow. User inputs are normalized into comparable fields, then surfaced through a simple compatibility layer so that people are not starting from scratch every time a new conversation begins.

Matching is treated as a scored process rather than a binary outcome. Practical constraints and softer preferences are evaluated together, allowing coordination to happen before conversation becomes unstructured.

Shared match views keep identifying details and direct contact information masked at the browsing stage. This allows users to explore compatibility without immediately bypassing the brokerage layer.

The system was designed not only for user convenience, but for brokerage-adjacent use in a real rental workflow. User-facing discovery and office-facing handling remain separate, so coordination, follow-up, and application support can still be structured.

A lightweight internal path is integrated into the same system. Agents can review submissions, inspect ownership-linked records, and move cases through defined states without direct backend access. The result is not a single matching feature, but a compact product system that holds intake, scoring, visibility control, and coordination together.

INTAKE MODEL

The first problem was not matching but comparability. Roommate search arrives as loose preference language, so the intake had to turn move-in timing, budget, geography, room type, roommate preference, lease length, pet tolerance, and optional compatibility signals into normalized fields without making the flow collapse under its own formality.

MATCHING LOGIC

The scoring model separates exclusion from ranking. Non-viable pairs drop out first. Softer preferences are weighted only inside the surviving pool. That avoids false precision while still producing a readable order of fit.

PRIVACY ROUTING

Each submission is assigned an internal identifier. Shared match views do not expose full identifying or contact detail. User-facing matching and office-facing handling remain intentionally separate, so the interface stays useful without becoming a direct bypass around the brokerage layer.

AGENT WORKFLOW

A lightweight front-end CRM path lets agents log in, review recent submissions, inspect ownership-linked records, and move cases through defined states without touching raw database access or backend tooling.

SYSTEM SHAPE

The build was end to end: product definition, interface design, matching logic, privacy handling, and workflow support. The deeper task was to make a fragmented, socially noisy housing search process legible enough to operate.

Live instrument

Data interpretation studio

A restrained prototype for reading messy tables. It checks structure first, then offers a small number of comparisons worth attention.

Open live demo

More work

Further entries from the same body of work. Shorter treatment here, but not secondary in importance.

02
Personal knowledge system
A maintained index for work, notes, references, and selected images — built as a long-form personal system rather than a portfolio shell.
This site is designed less as a showcase and more as a maintained personal system. Work, notes, archive material, and static identity pages sit inside separate sectional logics so different kinds of material can accumulate without flattening into the same presentation style.
The structure favors continuity over novelty. New entries are meant to be filed, placed, and revisited rather than endlessly pushed downward in feed form. Typography, spacing, navigation, and page rhythm are tuned so the site can hold analytical writing, visual fragments, project descriptions, and slower personal material without collapsing into a single undifferentiated voice.
The system is intentionally quiet. Instead of adding visible product behavior everywhere, small page-specific details are used to create recognition and continuity: restrained motion, margin objects, sectional pacing, and surfaces that feel editorial rather than app-like.
In that sense, the site functions as both interface and index: a place to present selected work, but also a framework for storing, extending, and shaping an ongoing body of material over time.
Personal system
03
Model stress-testing and evaluation design
STEM-grounded prompts, rubrics, and scoring structures built to make failure visible and comparable.
Quantitative and STEM-grounded evaluation tasks written to expose where strong models still break under constraint: hallucination under pressure, brittle reasoning, weak instruction-following, and instability when answers had to remain verifiable rather than merely fluent.
The work is adversarial only in a disciplined sense. Difficulty is used to separate surface fluency from sustained correctness, with ambiguity controlled so failures remain diagnostic. Rubrics keep those failures comparable across runs and tied to reproducibility.
Evaluation work

Selected work

Roommate matching system for shared housing

Data interpretation studio

More work