AI
Notes on building with AI, experiments I'm running, and the tools I use.
Projects
-
Ork
Task orchestrator built 100% with Claude Opus 4.5 using the Claude Agent SDK.
-
Twisted Timeline
Alternate history blog written entirely by AI. Used dynamic prompts with pseudo-randomness to generate new worlds daily.
Experiments
Three-Persona Code Review System
A sub-agent approach using three distinct reviewer personas, each running on Opus with structured investigation phases and formatted verdicts.
Bar Raiser — "Is this ambitious enough?"
Looks for missed abstraction opportunities. Pushes for changes that improve future work, not just current code. Verdict: PASS / ENHANCE / RETHINK
Devil's Advocate — "What could go wrong?"
Surfaces genuine risks and blind spots. Identifies silent breakage, hidden coupling, untested assumptions. Verdict: PASS / REVISE / BLOCK
Minimalist — "Can this be simpler?"
Looks for over-engineering and deletion opportunities. Favors straightforward over clever, deletion over refactoring. Verdict: PASS / SIMPLIFY / OVER-ENGINEERED
An orchestrating agent collects all three verdicts and determines whether to run another coding pass or continue to the next task.
Discovery
Tools
- Claude Code — Primary AI usage
- ChatGPT — Secondary AI usage
- 2080TI GPU — Running local models for experimentation
- Weaviate — Semantic search and RAG experiments
- Qdrant — High-performance vector similarity