Building AI Autopilot
Building an agentic loop for code, research, and workflows. Fewer assumptions. More control.
- Works on My Machine Is Not a Passing Test.
- Natural Language Was the Problem. Six XML Tags Were the Fix.
- Claude Gets Expensive. Codex Gets Killed. The Loop Doesn't Care.
- 300 Lines of Bash Replaced Six AI Skills
- BMAD Is Overkill. I Kept the Brief.
- 48 Lines. Still Wrong.
- I Built a Beautiful Machine and Then Deleted It
- I Was Feeding It Gaps. It Was Filling Them With Fiction.
- The Orchestrator Trap
- The Supervisor: Building the Machine That Watches
- Orchestrating the Swarm (Without Crashing)
- White Knuckles on the Wheel
- Now It Runs Faster Than I Can Think
- I Spent a Year Dragging the Machine
B2C Funnel Systems
Health funnels, paywalls, and the metrics that actually move money. No vanity. No blended scores.
- They Did Not Cancel. The Bank Did.
- How I'd Start Today If I Had to Start From Zero.
- Personalization Worked. Then Overfit.
- Good Ads. Bad Users.
- Cheap CAC. Expensive Business.
- Day-0 Lies. Day-7 Decides.
- Dashboard Green. Funnel Red.
- Screen Order Beats Offer Copy.
- One Paywall. Two Outcomes.
- Free Trial. Paid Regret.
- Quiz to Paywall. Where Money Dies.
No Bullshit Pipeline
Local-first recording and transcription on macOS. No cloud. No theater. Just a working pipeline.
- You Own the Pipeline. Or You Don't.
- Same Pipeline. Different Output.
- The Review Passed. Everything Was Broken.
- Transcription Is Step One.
- Building Desktop Apps That Don't Suck
- Whisper on Device: AI Without the Cloud
- Files Beat Databases for Personal Tools
- Recording macOS Without Root Access
- I Built a Spy Tool That Can't Spy
User Intent Prediction
How prediction models looked smart, failed hard, and got rebuilt from signal up.
- The Dumb One Won.
- Same Pattern. Different People.
- Thirty Days. Walking Dead.
- The More They Clicked, The Less Was Clear.
- Same User. Opposite Predictions.
- One Month Passed. Models Became Strangers.
- Position 40. Everything Collapsed.
- Calibration Fix.
- Doing Nothing Won.
- Smart Discounts Regret.
- Model Export Conversion.
- Universal Patterns.
- Time-Series Motivation. Model Said No.
- Two Tasks at Once. Forced Learning. Still Failed.
- Backward and Forward. More Work. Worse Results.
- Three Sequence Models. All Failed.
- 3,700x Faster. Still Not Fast Enough.
- Perfect Clustering Becomes Zero Prediction.
- 99.5% Accuracy. Zero Usefulness.
- How Does One Number Defeat a Neural Network?
- 32 Dimensions to 8. Everything Broke.
- The Plan