WRITING / POINT OF VIEW
Ideas for making enterprise AI actually work.
Verified investigations and long-form arguments on architecture, interoperability, operating models, governance, evaluation, and economics.
Your AI Agent Can Be Probabilistic. Its Guardrails Cannot Be.
Why “rogue AI” is the wrong lesson from the July 2026 OpenAI–Hugging Face incident—and how deterministic external guardrails can keep agents useful and contained.
Read the investigation →Architecture, control & value.
The AI Framework Compression Thesis
As interfaces standardize, generic framework value compresses and differentiation migrates upward into orchestration, governance, evaluation, and domain intelligence.
Read the thesis → Governance framework ↗Nine layers for useful, bounded agents
A practical enterprise control model spanning the task contract, tool gateway, identity, containment, approvals, monitoring, and lifecycle assurance.
Open the framework → Operating notes ↗Signals from building enterprise AI
Concise observations on retry tax, agent sprawl, context inflation, abstraction tax, and evaluation.
Browse field notes →Architecture over hype
Separate protocols, frameworks, models, and business operating controls instead of treating “agentic AI” as one layer.
Production over demos
Ask whether a system can be governed, evaluated, observed, maintained, and economically justified after the prototype.
Outcome over API call
Measure cost, quality, and reliability at the level of the completed task—not only the individual model invocation.