Essays
The Journal253 essays
Long-form writing on AI, philosophy, psychology, and systems thinking.
-
Designing Systems That Survive Their Architects
Systems with comprehensive documentation kept their operational effectiveness after the architect left. Systems built on tribal knowledge did not.
-
Writing as Thinking: Why Technical Leaders Write
Leaders who write weekly produce fewer dependency conflicts and fewer reversed decisions, because writing forces them to face objections first. Writing is the thinking, not documentation of it.
-
Database Architecture Decisions Outlast Every Other Technical Choice
In the systems I have built, database decisions made in the first months stayed the tightest technical constraint for years. Rare migrations were costly.
-
Evaluating LLMs Is the Hardest Problem in AI Engineering
Teams without systematic evaluation ship features with far higher defect rates. Evaluation is the hardest problem in AI engineering. Building rigorous evaluation infrastructure is the hardest and most valuable AI engineering problem.
-
The Ship of Theseus and Continuous Deployment
A continuously deployed application replaces nearly all of its code over time. The Ship of Theseus asks whether it is still the same system. The Ship of Theseus reveals that system identity is narrative continuity and purpose, not material composition.
-
Agile’s Midlife Crisis: Returning to First Principles
Teams closest to Agile's original principles had the fewest certifications and lightest processes. The industry built around Agile inverted its promise.
-
Observability for Ethical Systems: Monitoring Beyond Uptime
Adding fairness metrics to a large recommendation system revealed systematic biases that traditional uptime and latency monitoring had missed entirely.
-
Autonomous Agents Need Epistemology, Not Parameters
The bottleneck in autonomous agents is epistemic competence, not capability. Implementing uncertainty quantification markedly reduced confidently wrong actions.
-
Designing for Composability: Building Systems From Interchangeable Parts
Composable systems shortened feature development and cut integration effort for new channels where adopted, because new work reused existing parts.
-
Data Observability Needs Product Thinking
Standard observability catches most structural data failures but very few semantic ones. The semantic failures it misses cause far more business impact.