I build AI systems that hold up in real use — and the evaluations that prove it.

GG Tank Watch is a civic emergency archive I built during a real May 2026 chemical-tank evacuation (~50,000 residents; Wikipedia, NPR). While it ran, an LLM (Claude, with web search) summarized official and news updates every 30 minutes. A consumer-facing AI system with guardrails enforced by code and tests, not prompting:

• What if the AI got something wrong? The design assumed it would. The AI only wrote drafts — a separate, tested program checked every one and controlled what reached the page. And the page never told people what to do: at most, it pointed readers to official sources.
• The asymmetry that matters. A false all-clear is catastrophic; a false alarm is survivable. So danger downgrades need ≥2 sources (including an official agency), while upgrades fire on one. Enforced in code, never asked of a model.
• A behavioral harness of 200+ tests — grown from a red-team analysis of 12 real failure modes, green in CI — catches drift from the safety contract (fabricated sources, authored directives, stale data) before it ships, not after.

Start here (main repo) → gg-tank-watch

Live Website → ggtankwatch.org

Safety testing & evidence → gg-tank-watch-method

Business Scenario Judge — a method study: can you trust an AI judge to grade another AI? I blind-graded 60 real service chats in a labeling tool I built, closely calibrated an AI judge against a human expert's rulings — mine — and measured it against an uncalibrated judge on held-out chats — every ruling committed, all the math re-runnable by anyone, the honest result up front.

Three essays on AI evaluation, from practice — mikeilog.com/writing:

• What building an emergency dashboard taught me about trustworthy AI — the GG Tank Watch story: five lessons about AI you can trust.
• What AAA QA teaches about AI evaluation — Square Enix QA as a multi-agent evaluation system before anyone called it that.
• Human-in-the-loop eval and the part you can't ask about — what a real human-in-the-loop looks like, and when the honest move is to cut the feature.

Open

• gstack security fix #1822: withholds the localhost auth token from content-script getPort callers
• Claude Code
  • Reported and diagnosed session-bloat bug #61613 in the Read tool's binary-file serialization path (root cause + fix sketch)
  • Added a reproduction, render-only proof, and fix direction to scrollback-duplication #51828

Merged

• gstack PR #1554: fix shipped in squash commit 7ca04d8, credited in the v1.42.0.0 release wave (#1594)

• Building AI agents: Claude Agent SDK, MCP, Anthropic SDK
• Languages: Python, TypeScript

• LinkedIn → https://www.linkedin.com/in/mikeilog
• Email → mike@mikeilog.com

_{cooperation FTW · US (Pacific time) · remote}