Case study 3

The AI Airlock

Spend the AI tokens once to build; let plain code do the daily work.

  • Architect
  • Product
  • Build

Project record

Role
Architect · Product · Build
When
Aug 2026 to today
Status
Live inside the Online Employment Center
Built with
Python, PostgreSQL, a manifest-based load step, Claude via a controlled hand-off

Next: The problem ›

The problem

Many teams now send most of their work to AI agents. That burns tokens every day, costs keep rising, and an agent with access to production can do real damage.

Next: What I did ›

What I did

I used AI to help me build, then designed the products to run mechanically. Around 90% of the daily work is handled by predictable code with no AI at all. The small part that benefits from AI goes through the Airlock: data out through one door, results back through another, checked by a person before anything is loaded.

Next: What shipped ›

What shipped

A working Airlock that handles the Career Bio and the final fit check in my Job Bank, with no agent in production.

Next: What I learned ›

What I learned

Treating AI as a narrow, deliberate layer makes a product more cost-effective, safer and far easier to keep compliant.

Next: Open it live ›