SOFTWARE / SYSTEMS / ELECTRONICS

Curiosity,
made real.

I’m Yingche — a student engineer and ScholarSafe co-founder. I build across AI, software, and electronics, following an idea all the way to something I can test.

BASED IN SAN JOSE, CASTILL ASKING “WHAT IF?”
THE WORKBENCH 01—04
Student housing, roommates, and the trust between them.Explore
SELECT A BUILD. FOLLOW THE CONNECTION.

Different projects.
The same instinct.

Understand the system.Build the real path.Test what breaks.

01 / SELECTED WORK

The work I’d
start with.

Product ownership. Systems depth. Applied AI. And hardware that leaves the screen.

01PRODUCT / CO-FOUNDER

A place to live.
People to build with.

ScholarSafe ↗

Housing search, roommate matching, and the trust workflows around them. I co-founded ScholarSafe and lead the engineering behind the student experience.

Authentication, weighted matching, messaging, verification, and moderation. Built around an actual student problem.

Actual ScholarSafe homepage showing student housing discovery and an explicitly labeled example listing
Actual product interface · example housing and roommate data
02LINUX / DISTRIBUTED SYSTEMS

The interesting part
is what survives.

HomeHub ↗

A local AI and automation system connecting an HP Envy core and a Raspberry Pi edge node. Durable state, bounded workers, and explicit recovery turn a collection of services into a system.

A timeout does not grant permission to replay work. A green build does not grant permission to rotate production.

2NODESOne explicit protocol. Local SQLite state.
Core 1.7.4 · advanced execution remains gated
HOMEHUB / CORE + EDGEARCHITECTURE WALKTHROUGH
EDGE NODERaspberry PiBounded edge services
CORE NODEHP EnvyState, policy, workers
LOCAL STATESQLiteSERVER-OWNED POLICY
01 / BOUNDED EXECUTION

A worker gets a boundary.

Sessions, leases, and cancellation are owned by the server. Tool access stays bounded.

Private infrastructure, explained here. No connection to the home network.

Ω / OHMNIEXPLORE A RESISTOR MIX-UP
SCHEMATIC EXPECTS2.500 V
ILLUSTRATED OUTPUT2.500 V
A five-volt divider. R1 is 10 kiloohms; toggle R2 between 10 kiloohms and 1 kiloohm. VOUT is measured at their junction relative to ground. 5 V R1 · 10 kΩR2 · 10 kΩ VOUTDC supply

Two 10 kΩ resistors divide the 5 V supply equally.

Calculated illustration: VOUT = 5 × R2 / (R1 + R2). This is not a live measurement.

03CIRCUITS / APPLIED AI

The drawing says one thing.
The meter says another.

Ohmni ↗

A schematic becomes a reviewable graph. A deterministic solver does the math. Your measurements turn a mismatch into testable fault hypotheses.

Models help read the circuit. You confirm the connections. The solver owns the numbers.

04LOCAL AI / VERIFIED SKILLS

Solve it once.
Remember how.

SkillForge ↗

A local AI runtime that turns a verified, human-approved fix into a reusable skill. The next compatible incident runs the checked code with zero model calls.

9/9hidden checks passed by a model-authored skill
Recorded GB10 evidence · synthetic incidents
MEET EMBERINTERACTIVE EXPLANATION
Ember, SkillForge's original pixel companion
01 / INVESTIGATE

A new problem. A little spark.

The local model reads the incident and proposes a bounded repair.

An explanation of the system. No live model or backend calls.

05RUST / EMBEDDED LINUX / FPGA

Some ideas need
a power switch.

Cyberdeck ↗

A Raspberry Pi, a touchscreen, and a Rust interface. Underneath: local status messages, a control daemon, and an FPGA cooling core tested in simulation.

60FPSRust / Slint on a Pi Zero 2 W
From the physical validation record
CYBERDECK / V2ARCHITECTURE STUDY
CYBERDECKRUST / SLINT
Built to run
close to the metal.
UILOCAL IPCCONTROL
YW / 02

Rust / Slint renders directly through LinuxKMS. Touch navigation and boot recovery are physically validated.

Device illustration, not a photograph. Physical fan integration is pending.

02 / THE REST OF THE WORKBENCH

Always something
on the bench.

Products, infrastructure, experiments, and unfinished questions. Every build has a different stage.

13 builds in the index

01

ScholarSafe

Student housing discovery, compatible roommates, and trust workflows.

PILOT PRODUCTPRODUCTS
02

HomeHub

Durable local state, bounded workers, and explicit release controls.

PRIVATE SYSTEM · CORE 1.7.4SYSTEMS
03

Ohmni

Schematic to solver-backed circuit diagnosis.

WORKING PROTOTYPE · DEVELOPMENT CONTINUESAGENTS
04

SkillForge

Verified local AI skills. Reuse a checked fix without another model call.

WORKING PROTOTYPE · RECORDED PUBLIC DEMOAGENTS
05

Cyberdeck

Rust, a Raspberry Pi, and a simulation-tested FPGA cooling core.

HARDWARE VALIDATED · FPGA IN SIMULATIONSYSTEMS
06

DayBridge

Activity planning that accounts for practical barriers and consent.

HACKATHON PROTOTYPE · SIMULATED PROVIDERSPRODUCTS
07

Converge

Recover partial workflows without repeating completed side effects.

HACKATHON PROTOTYPE · HOSTED DEMO TEMPORARILY OFFLINEAGENTS
08

ColleAgent

Private student planning meets public academic evidence.

PRE-PRODUCTION ALPHA / PILOTAGENTS
09

After the Shield

What happens to a robot policy after a safety intervention?

RESEARCH / EXPERIMENT SETUPRESEARCH
10

Internship Finder

Evidence, eligibility, and explainable opportunity ranking.

OWNER LOGIN REQUIRED · PUBLIC SOURCEPRODUCTS
12

Slopebot

Canvas perception, feedback control, and repeatable browser experiments.

EXPERIMENTAL · FIXED 960×540 LAYOUTSYSTEMS

03 / THE PERSON AT THE BENCH

Engineer in progress.
Curious by default.

I’m a dual-enrollment student at Mission Early College High School and Mission College, in San Jose, California. My projects move between software and the physical world because I like understanding the whole system.

Outside the builds, I founded my school’s AI Club, led 12+ Python and AI workshops, and collaborate on educational AI evaluation with San José State University.

Education, experience & the longer story
A FEW THINGS I KEEP COMING BACK TO
01

Build something useful.Start with a person and a real constraint.

02

Make the uncertainty visible.A hypothesis, a simulation, and a result deserve different labels.

03

Try the thing that might break.Recovery and failure are part of the design.

04

Leave a trail someone can follow.Source, evidence, and decisions belong beside the work.

THE NEXT CONNECTION

Have a good
“what if”?

Let’s build something.

SKILLFORGE

The Ember workspace.

SkillForge Ember development dashboard with the active mascot, investigation map, workflow steps, and recorded GB10 evidence

Development dashboard preview. The public guided demo uses recorded and synthetic data.

Open the public guided demo ↗