I build agent systems, interactive learning software, and experimental interfaces — from an on-device Android agent and verifiable computer automation to tools that make knowledge more accessible and learning more active.
A selection of current software, research, and hardware projects. Public source is linked where available; private prototypes and early-stage research are labeled accordingly.
An autonomous agent designed to live on an Android phone. Umbra combines accessibility-based app control, optional privileged execution, a local Linux development environment, and background task sessions. Its execution model emphasizes explicit approval for sensitive actions, verification, resumability, and keeping the user in control.
A learning workspace for IB-oriented study materials, collaborative notes, interactive lessons, native quizzes, flashcards, and practice examinations. I am exploring how AI-assisted content generation can connect directly to reusable learning interactions instead of stopping at static summaries.
Local Human Intent Controller is a local-first browser and desktop automation runtime. A deterministic, model-free Fast Path handles supported actions; an optional model-guided Slow Path proposes steps that remain subject to validation, human approval, verification, and audit trails.
An exploratory research direction on how an agent can understand and operate structured graphical interfaces with less dependence on fragile pixel-level control. Current work focuses on synthetic UI environments, representations, controlled ablations, and checking novelty against prior research — not on claiming a finished method.
A programmable 84-key mechanical keyboard concept built around an ESP32-S3, a full-width touch display, additional status screen, macro keys, and physical controls. The work includes CAD, component-fit studies, enclosure design, and a dual-device connectivity plan; hardware fabrication is still in progress.
A guest-connectivity and ordering infrastructure concept that links POS-driven table sessions to short-lived, revocable network access. Current engineering focuses on access lifecycle controls, authorization boundaries, simulated integration flows, and partner-facing validation before real-world deployment.
Project maturity varies. Linked repositories are public; other work is private, experimental, or not yet released.
The Interactive Classroom Compiler
IdeaStone is an in-development learning platform centered on compiling source material into structured, reusable interactive classrooms. The goal is to combine adaptive tutoring, a learner model, and verifiable learning evidence while keeping high-quality courses affordable. The public product is not being presented here as a completed launch.
Assemble lessons from reusable explanations, simulations, exercises, checks for understanding, and learner-controlled interactions rather than one-way generated text or video alone.
Explore structured course compilation, source-grounded quality checks, tutoring workflows, and learner-state tracking that can guide the next useful activity without obscuring what a student actually understands.
A core design goal is inexpensive, precompiled public lessons with optional AI assistance, so interactive education can reach more learners without making every activity depend on expensive inference.
My work moves between research and implementation: representations, execution, evaluation, user interaction, and the constraints that decide whether a system can work outside a demonstration.
How can autonomous actions across phones, desktops, and applications stay verifiable, approval-aware, and recoverable?
How can interactive instruction adapt to evidence of understanding instead of relying on passive content or guesswork?
Can structured application capabilities replace fragile UI control with efficient, deterministic agent-accessible operations?
How do security boundaries, constrained hardware, reproducible tests, and deployment realities shape better architectures?
I am still early in my academic path, so this site emphasizes active research, technical work, and research direction rather than presenting unfinished work as publication.
I am Pinyen Cheng, a student researcher and engineer at Taoyuan Municipal Dayuan International Senior High School in Taiwan. I learn best by building systems that force me to understand the underlying problem.
My interests span on-device AI agents, adaptive learning, interface representations, embedded hardware, networking, developer tooling, and practical software engineering. I am particularly drawn to projects that are slightly too ambitious for the resources available, because constraints tend to expose the architecture that actually matters.
My long-term goal is to contribute to intelligent systems that are not only more capable, but also more efficient, interpretable, reliable, and useful in real-world settings.