Not every agent step deserves the largest model
What changed after treating agent traces as routing signals instead of passive logs.
Notes
A record of understanding in motion. Not a catalogue pretending everything has already been mastered.
An AI failure needs a name before it needs a dashboard
Why failure taxonomy became the foundation of useful evaluation and root-cause analysis.
The contract breaks between the services
What a cross-component test suite revealed that isolated unit tests could not.
Archive the data, not the latency
How treating retention as architecture reduced database pressure and improved an API by 60%.
AI was useful only after it fit the workflow
A production AI integration reduced a manual review from seven minutes to under two—but the model was only one part of the result.
The stated requirement is rarely the whole problem
What customer-facing delivery taught me about translating requests into systems that survive reality.