Vibe-Coded Evals with LLM-as-a-Judge

December 13, 2025 • 2 min read

Using Claude Code and OpenRouter infrastructure to rapidly build model evaluations with Claude Opus 4.5 as the judge

Whisper Cantonese Transcription: Force zh, Not yue

Counter-intuitive finding: forcing Cantonese (yue) on Whisper causes decoder collapse. Force Chinese (zh) instead for coherent, searchable transcripts—even for Cantonese content.

January 13, 2026

Why I Open-Sourced a Deterministic I-Ching Engine (Or: The Smallest Verifiable Compute You Can Build)

Extracting the deterministic core of 8-Bit Oracle into a standalone library, and why 'same seed, same hexagram, every time' matters more than you'd think for decentralized AI.

December 1, 2025

Testing Twelve AI Image Models with Marcel Proust's Madeleine (Or: How We Chose Hunyuan for Hexagram 24)

Upgrading 8-Bit Oracle's Hexagram 24: Return with Proust's famous involuntary memory scene, then systematically testing 12 diverse AI image models through fal.ai to find which one actually captures that precise moment when past and present collapse together.

November 18, 2025