All work, by area
AI products
Live products under Mind Theory AI.Live · ghostgraph.aiApr to Oct 2026
Ghostline: graph intelligence that labels every inferenceA screening and investigations platform that shows how people and entities connect, and never presents a guess as a fact. One product in the Ghostgraph.ai suite, shipped to production as 16 PRs in 11 days.16 PRsmerged to production in 11 days
Working locallyMay to Oct 2026
Khun: live Thai–English call translation with measured guardrailsA real-time translator for calls where speakers switch between Thai and English mid-sentence, running on a laptop.0.988 → 0.174English WER on code-switched audio
Live · lawwiki.aiDec 2025 to Oct 2026
LawWiki: California law in plain English, citation-checkedA public legal-research site with hybrid search and quote-level citation checks.126,650+California statutes, 29 codes
Applied AI and evaluation
Models chosen and gated by measurements that can't be faked.In production (local)Aug to Oct 2026
Transcribing near-silent audio without inventing wordsA local transcription stack that holds the leaderboard's best model back until an independent model confirms there is speech.93.5%of no-speech clips got Whisper text
Daily driverOct 7 to 9, 2026
Claude Code on a laptop-local modelRan Anthropic's coding agent against an open-weights model on an M4 Max, with 83% fewer prompt tokens per request.83% fewerprompt tokens per request
GTM engineering
Signal-based pipelines with verification built in.Data and reliability engineering
Verify first, then act. Every gate proven able to fail.Lab
- Trend PulseLiveGitHub's top 100 repositories over 7, 30 and 90 days, with true watcher counts, refreshed by a scheduled GitHub Action.Next.js 14, Python fetchers, GitHub ActionsOpen
- Ikigai QuizLiveA career-fit quiz built on four established psychology frameworks.Next.js 14, zustand, zodOpen
- Call audio-route monitorPrototypeA macOS monitor for per-process audio routing during calls, with hash-chained logs. No root and no microphone access.Python (stdlib + ctypes into CoreAudio)