Posts
Posts about software, AI, and building things.
When Your Agent's Alarm Clock Breaks
My morning briefing arrived at 10 PM. The root cause was daylight saving time, a UTC date boundary, and a scheduler that doesn't handle the intersection well.
Giving Claude Code a Memory
Claude Code sessions are stateless. Every conversation starts from zero. Here's how I built a persistent memory layer so each session can search everything that came before it.
The Token Budget Problem Nobody Designed For
Enterprise AI spend shifted from training to inference. The constraint isn't compute — it's that nobody built agents with token budgets in mind.
Writing Postmortems for Your AI Agent
I've written two formal postmortems for an AI assistant in a single week. Here's why that's not weird.
What Model Degradation Actually Looks Like
Context saturation doesn't crash a model. It changes its personality. Here's what that looks like in production.
Three Models, One Agent
Running Gemini Flash, gpt-5-mini, and Sonnet 4.5 in the same system taught me that model selection is an operational decision, not a technical one.
Prefetch Changed Everything
How moving data collection out of the AI loop cut token costs by 99% and turned a fragile demo into a reliable system.
Building an AI Email Triage System
Why I built an AI agent to monitor email across multiple accounts, and what the first version looked like before I learned anything.
Cron Isn't Sexy But It's Load-Bearing
When every cron job is an API call that costs money, scheduling stops being a deployment detail and becomes architecture.
Running an AI Assistant on a $300 Mini PC
A Beelink SER5 runs my entire AI automation stack. Here's what that looks like and why it works.
Tools I Actually Use
Not a recommendation list. A field inventory of what's installed, what gets opened every day, and what's gathering dust.