Hypotheses should be testable.
Architectures should be falsifiable.
Tradeoffs should be explicit.
Evidence should outrank convention.
If the evidence changes, so do the opinions.
I build and examine AI systems through that lens: as claims to test, not fashions to adopt.
- AI architecture matters more than model choice.
- Governance belongs in the architecture, not the policy binder.
- Most AI failures are systems failures, not model failures.
- Systems fail more often at boundaries than at components.
- Every AI decision needs an economic justification.
- Simplicity compounds.
- Evidence beats convention.
Question: Can agentic systems remain governable without sacrificing capability?
Current evidence: Safety belongs in the wiring, not the model.
Question: When is GenAI worth 150× the cost of SQL?
Current evidence: Far less often than marketing suggests. The architecture should prove when AI is justified.
Question: Can a thirty-year-old system be modernized without rewriting it?
Current evidence: Architectural seams age better than rewrites.
Question: Can architectural reasoning become reusable?
Current evidence: Good workflows preserve judgment, not just prompts.
I used to believe frameworks accelerated AI development.
Now I believe architectural boundaries matter more.
I used to believe better models solved most engineering problems.
Now I believe orchestration, constraints, and evaluation matter more.
I used to believe retrieval was the hard part.
Now I believe authorization, provenance, and governance are harder.
Questions worth testing:
- What if governance is a compiler problem?
- Are identity graphs better understood as probabilistic databases?
- Does every architecture diagram lie by omission?
- Is evaluation the missing interface in most agent systems?
- How much complexity exists only because the boundaries are wrong?
- When is AI the wrong answer?
Ideas that survive become articles.
Articles that survive become reference architectures.
I write about architecture, AI governance, systems thinking, and the economics of technical decisions.
Hypothesis: Skynet Hunter should run against AWS Bedrock without changing the core architecture.
Only the model adapter should change.
If that proves false, the system is more tightly coupled than the diagram admits.
Shaped in the Midwest. Refined in Berlin. Still testing hypotheses.