Dev-log

A deliberation session silently failed because the health check was hitting the wrong endpoint. Fixing that one bug turned into a five-part architecture overhaul covering provider dispatch, transcript persistence, and orchestrator handoff.

How I built a multi-provider dispatch router that sends different AI personas to different LLM providers and the $57 billing surprise that proved it was worth it.

How I built a system where three AI providers debate feature decisions before any code is written, and why the disagreements produce better designs than any single model.

How I directed a circuit breaker feature across llm-router and AutoAgent using an AI development pipeline, and what the automated review caught before it hit production.

EcoOrchestra governs every step. Ralph ships every loop. Both work. The question is which failure mode you fear more.

Ghostpen turns development artifacts into blog posts written in my voice -- because the best time to write about a problem is when you just solved it.

An AI agent that learns to play games by watching VODs -- extracting knowledge with vision APIs, compiling behavior scripts, and executing them with computer vision.

A multi-provider LLM routing library that handles rate limits, batching, and failover -- so consuming applications never touch provider-specific APIs.

How I built an AI development pipeline that runs a team of personas -- planner, developer, reviewer, QA engineer -- and enforces quality gates across multiple repos.