The three failure modes that kill personal AI agents
Every failed personal AI assistant project I've seen breaks down the same three ways. Here's what they are and how to avoid them.
Read →BLOG · PAGE 5
Page 5
Two scoring systems that never talked to each other
The job pipeline scraped 64 fresh listings and the apply cron found zero qualifying ones. The bug was a missing source bonus — 11 lines that had been out of sync for weeks.
Read →Eight products, one night: what autonomous AI agents actually ship
I've been running Claude autonomously on my homelab from midnight to 6 AM. Here's an honest accounting of what it's actually produced.
Read →Real-time voice triage with Twilio, Claude, and ConnectWise
How NightDesk's real-time triage kernel routes after-hours MSP calls — every turn of the conversation, not after the call ends.
Read →The 6-hour window: why you need to know when your product appears on HN
A Hacker News thread about your product is the highest-leverage 30 minutes of your day — if you're there for it. Most founders aren't.
Read →I automated my job search. Here's what actually works.
Six months of building a job application pipeline with Claude and AWS. The architecture, the failure modes, the current state of the system.
Read →When the pool runs dry: scoring functions and the empty-pipeline problem
My auto-apply pipeline hit a wall: 971 job listings marked new, zero qualifying applications. Here's what the scoring distribution revealed and what to do about it.
Read →Why TicketScope is a Chrome extension first, not a Marketplace app
The ConnectWise Manage app marketplace would take 6-8 weeks to review. A Chrome extension takes 60 seconds to install. Here's why I started with the extension.
Read →The SQS FIFO queue that lets a Lambda run 50 overnight jobs without timing out
Lambda has a 15-minute execution limit. An overnight job that processes 50 items needs hours. The coordinator/worker/SQS FIFO pattern solves this — each item gets its own Lambda invocation and full timeout.
Read →The one wrapper that makes LLM tool calls production-ready
An LLM agent that lets exceptions propagate from tool calls will silently stop working at 2 AM. One wrapper function prevents this — and it changes how you design every tool.
Read →The OSS dependency that was unmaintained for two years before we noticed
A production service depended on a library whose last commit was 2019. No CVEs, no alerts — the library just quietly stopped being maintained. We noticed when it blocked a Node upgrade.
Read →What happens when your MSP client calls at 11pm
Most MSPs have an after-hours number that goes to a human. One call center quote later, that answer becomes 'voicemail.' Here's what I built instead.
Read →The resume that updates itself
Why resumes are always stale, and the GitHub-connected CV that patches itself overnight from your actual commit history.
Read →The DynamoDB memory that survived 6 months of production
How to give a personal AI agent durable memory using DynamoDB — the four-table schema that handles conversations, facts, preferences, and tools without ever running a migration.
Read →