IanYu

Building systems that let one person think, create, and learn at greater scale.

LIVE SYSTEM FIELDPOINTER / MOTION

Products feed agents. Agents compound knowledge. Knowledge informs capital. Capital buys time to build again.

CURRENT INQUIRY

What can one person build when products, agents, knowledge, and capital share the same operating system?

00 / SYSTEMLOOP, NOT MENU

Four forms of leverage. Not four services.

They form a loop: software turns ideas into tools; agents expand execution; knowledge preserves judgment; capital buys back time to build again.

01ProductsIdeas become software
02AgentsSoftware learns to coordinate
03KnowledgeWork becomes memory
04CapitalMemory informs risk

01 / PUBLIC ARTIFACTS02 ENTRIES

Not a portfolio. A trail of verifiable work.

Every artifact begins with a question and leaves behind an intervention, technical decisions, evidence, and sources.

ARTIFACT 001Product / 2026v2.0.0

SpeakOnly

A low-latency voice input tool designed for Traditional Chinese users in Taiwan, with mixed Chinese-English transcription, custom terminology, spoken-language cleanup, automatic formatting, and optional on-device processing.

QUESTION: How can voice input for Traditional Chinese in Taiwan combine low latency, deep customization, and data ownership?
INTERVENTION: Reworked the byetype performance path to reduce 4–6 second latency to hundreds of milliseconds, then added offline local models, Taiwan-specific OpenCC vocabulary, Markdown prompts, history controls, a new brand, and AI image text extraction.
EVIDENCE: v2.0.0
4–6 s → sub-second
~8 MB app
macOS + Windows
STACK: Tauri 2 / React 19 / TypeScript / Rust / macOS / Windows / Local AI
SOURCE: devonmochi/byetype / MIT

ARTIFACT 002Product / 2026COMPLETED

AI Desktop Assistant

A local, full-duplex multimodal desktop assistant for macOS and Apple Silicon that can perceive screen and system audio while keeping its core inference workflow on-device.

QUESTION: Can a local multimodal AI system designed around Linux and NVIDIA be rebuilt as a complete native macOS workflow?
INTERVENTION: Independently ported the upstream Windows architecture to macOS, integrating ScreenCaptureKit, Metal inference, Unix domain socket IPC, the Python backend, and Electron packaging into a verified end-to-end system.
EVIDENCE: 8 phases completed
0.51 s average multimodal prefill
STACK: macOS / Apple Silicon / C++20 / Python / Electron / Metal / ScreenCaptureKit
SOURCE: LYiHub/pub-local-jarvis / MIT

02 / FIELD LOGCONTINUOUS

The working system has versions.

The log tracks changes to research methods, tools, and human-agent collaboration—not just website releases.

v0.4.02026-08-05

The site becomes an artifact

Reframed the homepage from a product landing page into a hybrid digital studio: an editorial field archive with an interactive system map and evidence-led artifact dossiers.

03 / NOTES / REPORTSOPEN RECORD

Make the process legible, not just the outcome.

Obsidian preserves long-term context; Ian’s Reports publishes research. Agents may collaborate, but judgment and accountability remain with Ian.