Work
Built end to end, measured in production
The featured cases below started as real problems, shipped as working products, and are still used. They are the evidence behind every service on this site.
Personal RAG Knowledge Base
A production-ready personal RAG system over 42,000+ knowledge documents - multi-workspace semantic search (LL work, MindX consulting, personal, shared research, AI vendor canon) accessible from Claude Desktop, Claude.ai web, and iOS app via the Model Context Protocol.
Read the caseMCP Connector Hub
One MCP server that gives Claude access to Google Workspace, Slack, GitHub, Jira, Confluence and Linear across three separate workspaces, with each workspace's data kept apart by design.
Read the caseEval-Framework
A personal eval harness that turns LLM model + prompt + scorer decisions from gut-feel guessing into evidence-based answers via stratified eval sets, multi-provider bake-offs, and LLM-as-judge scoring.
Read the caseKnowledge-Audit
A 3-layer audit + 4-tier cron that catches stale facts and cross-source contradictions in a personal RAG knowledge base - caught 25+ drifted facts in 79 files on day-one deploy.
Read the caseMail-Assistant
A local-first AI inbox triage layer that classifies threads P0/P1/P2/NOISE with cross-channel verification - cutting daily triage from 90 minutes to 8 minutes.
Read the caseDiagram-Engine
A text-to-diagram DSL whose primary user is an LLM, not a human - token-efficient syntax, deterministic SVG layout, and validator errors that LLMs can repair in one shot.
Read the caseMore work
Smaller tools, experiments and one concept that has not been built yet.
Auditing and rebuilding vuihoc.ai against Apple's HIG
How this website was audited against Apple's Human Interface Guidelines and WCAG with measured facts, then rebuilt from a personal blog into a services site.
AI-Canon-Crawler
A sister tool to Personal-RAG that crawls authoritative AI vendor docs into a dedicated _canon workspace, with routing rules that prefer vendor truth for spec/price/version questions.
Voice-Assistant
Voice-first agentic assistant running native on macOS - push-to-talk or continuous mode, controls 40+ Mac actions (open apps, type text, click, browser tabs), searches personal KB, multi-LLM (Cloud Haiku / Local Hermes / Apple Foundation Models). Pure Swift, ~50MB DMG, no Electron.
Health Coach
A personal health coach over Garmin data lake with ADHD-friendly nudges. Daily morning ping + Saturday digest + Calendar pre-meeting breath alerts via Telegram. Goal - weekly avg sleep score ≥80.
LumiConcept
The first conversational SMS for SEA preschools - teachers narrate the child's day by voice, principals run the school through chat. AI drafts, humans approve, the system executes.
Mac-Translator
A single-purpose macOS menubar translator triggered by Force Click or ⌘⌥T - selection becomes a floating HUD next to the cursor, no window, no flow break.