Essays on applied machine learning, cloud architecture, and the product-design instinct that should sit underneath both. Long-form, occasional, opinionated.
The non-human identity crisis is measurable in your own account today. Here's a concrete audit — admin agents, secret age, secrets outside repos — and what to fix.
The non-human identity crisis is measurable in your own account today. Here's a concrete audit — admin agents, secret age, secrets outside repos — and what to fix.
No agent should ever hold a standing key. AWS STS and OAuth token exchange already make that possible — the failure is architectural, not technological.
Most AI features don't fail because the model is wrong. They fail because the product never earned the user's confidence — and confidence is something you design, not something you ship by accident.
A guardrail in a system prompt is advisory. Authorization at the MCP gateway is enforced before the tool call runs. Here's the difference, on AWS.
Everyone's bracing for prompt injection. The boring, bigger risk is that your AI agents are non-human identities multiplying 144:1 — and most are holding long-lived secrets.
A field guide to the SageMaker patterns that survive contact with production — and the well-architected guardrails that keep them there.
Showing 6 of 7 — older posts paginate here once the archive grows.
A low-volume letter on applied AI, architecture, and the craft of building products. Double opt-in via ConvertKit — no spam, unsubscribe anytime.