Original research, project retrospectives, and unfiltered takes on AI in practice.
February 11, 2026
What happens when an AI agent starts optimizing its own infrastructure? A real investigation into emergent reward hacking.
AI-generated code is disposable by design. The real asset is the specification that generates it — and the process that verifies the output.
Building the world’s first multi-tenant MTA platform. KumoMTA is the engine — CompleteSender is the vehicle.
Infrastructure orchestration that actually configures production systems — not just deploys containers and hopes for the best.
Why every AI output needs a hostile reviewer. The case for adversarial quality gates that actually block bad output.
The RapidLaunch orchestrator was built in Python over 18 months. It worked. We rewrote it in Go anyway. Here’s why.
Stop asking developers “is this for production?” and start assuming it might be. Foundation Standard as the new default.
Python’s dynamic typing makes AI-generated code fundamentally harder to verify. The training data feedback loop is making it worse.
The full technical argument for why statically typed languages produce better AI-generated code. 25 minutes of evidence.
Subscribe to get new investigations and technical retrospectives. No spam, no fluff.
Complimentary 30-minute technical assessment. No commitments.