AI engineering
Shipping agents and LLM systems that hold up outside the demo: grounding, evaluation, and the failure modes nobody screenshots.
MCP or a direct API? Wiring tools into agents
The same tool built twice, and a rule of thumb for choosing between a direct tool call and an MCP server.
Constitutional AI without the philosophy
A working LLM-judge loop with a bounded retry budget and a visible audit trail when it fails to converge.
What today's AI actually costs
The bill moved left, from running software to experimenting with it, steering it, and evaluating it. A field report on where the money really goes, and why so much of it never ships.
When to abandon the low-code path
Why I rebuilt my procurement agent from scratch in week three of a three-month internship.
Anti-hallucination through tool grounding
A small experiment in forcing every claim from an LLM agent through a deterministic database query.
Notes on AI-assisted development
Patterns that have held up across six months of work with Claude Code.