7 posts
Training makes the headlines. Inference is what actually shows up on the bill, every request, forever, and it now eats over 80% of AI GPU spend.
A malicious model doesn't look wrong when you read it. Real cases of poisoned datasets, backdoored models, and compromised MCP servers, and how to actually verify what you're running.
The demo always works. RAND found over 80% of AI projects still fail, and MIT found 95% of pilots never show up on the P&L. Here's exactly where they die.
Reading five good-looking outputs isn't evaluation, it's a guess with extra steps. Real metrics, golden datasets, LLM-as-judge, and a CI/CD gate that actually works.
'Application plus LLM' was never the architecture, it was the demo. Here's the eight-layer stack a real AI system actually needs, and a map to every layer.
A 200 OK and a wrong answer. Traditional logs can't tell you why an agent decided what it did. Here's what to actually record, and how to trace it back.
The model call is the easy 5%. Here's the architecture, security, cost, monitoring, and reliability work that turns a demo into a system you can trust.
