Blog

Why AI projects in banking stall at the compliance boundary, not the model, and how boundary-first architecture ships under DORA and the EU AI Act.

When an AI agent ships code, who is accountable? A sign-off means little if the AI could rewrite the evidence behind it. Two questions to ask first.

MiniCPM5-1B runs a smart home because a Rust harness treats it as untrusted input. What EdgeHome Harness owns, why frontier agents need the same.

Document parsing on n8n is commoditized. The value is the decision before the build: throughput threshold, build vs buy a parser, and self-hosting for compliance.

How I wrap a coding agent in a sandbox so each risky capability, from model routing to patch approval, sits behind its own boundary.

P2P local AI: models arrive like torrents, inference can delegate to peers, no cloud in the path. Tested on a 4-core VPS with no GPU.

Agents restart from zero every session. One half of that memory is derivable from your code, one half only humans can author. Mixing them wastes both.

AI agents fail decisions for the same four reasons people do. The WRAP framework, encoded as a system-prompt gate, fixes what a bigger model can't.

Your AI agent efficiently executes tasks but misses hidden dependencies between projects. Beads solves this with Git-backed dependency graphs that reveal your true critical path across all work.

Execution-only AI agents finish every task but never ask if the task was worth doing. The fix is a steering layer over your work.

Your AI agent reaches for googled best practices before your own proven fixes. Wire a trust order into your CLAUDE.md and agent loop instead.

AI coding agents agree to your process, then skip it. Why review can't catch it, and the one fix that works: a deterministic finish-line gate.