Field Notes

Every automation, shipped in the open.

Build logs and short lessons from running a fleet of AI agents solo. Every post is a real thing I built, broke, or fixed. Proof, not promises.

2026-09-07 Salesforce, cleaned The safe way to let a robot clean your Salesforce. Letting automation loose on your CRM is terrifying. Here's the method that makes it safe: read-only audit first, you approve the plan, every change logged and reversible. read → 2026-08-10 Build Notes I got tired of my AI forgetting me. So I built it a memory. JayBrain: the memory system I built from scratch so the assistant stops starting every day with amnesia. A real vector database, hybrid search, and the boring parts that make it trustworthy. read → 2026-08-06 The Solo AI Fleet I gave one person an AI army. Here's what broke first. The failures that cost me real time had nothing to do with the model. They were the boring infrastructure around it, and that's the whole lesson. read → 2026-02-22 Build Notes I built a watchman for my own machine. A Python daemon that watches my own systems and only speaks when something needs me. Ten scheduled checks, rate-limited alerts, and the Windows bug that spawned eight ghost processes. read → 2026-02-21 Security Lab I taught my lab to notice a lockpick. Portable detection rules, a SIEM timestamp bug that quietly broke every correlation, and attacker-technique-mapped alerts deployed against a Windows Active Directory lab. read → 2026-02-20 Security Lab I attacked my own network to catch the attack. I ran a real credential-theft attack against my own domain, watched what it left in the logs, and wrote the rule that catches it on a fifteen-minute clock. read → 2026-02-19 Security Lab I locked the doors before I hung the cameras. Credential hygiene and the first monitoring dashboard: the unglamorous, load-bearing work of securing the pipeline before anything interesting can be watched. read →

Soon Running AI Cheap Your "always-on AI employee" is just cron, git, and a cheap server. How one small scheduled job runs an entire project portfolio. No Kubernetes, no agent framework.

Soon AI You Can Trust I caught my AI inventing a number. Then I built a machine to block it. A specific with no source is a hallucination until proven otherwise. The small commit hook that enforces it.