AI Cloud Lab
Real-world AI + AWS experiments. Each lab ships with architecture, code, cost, security, and a write-up.
The 0-to-1 opener: a useful bilingual concierge that lives in a chat app, built as n8n + one Claude tool-use loop. Three production-judgment guardrails — a hard cost-cap, retries with idempotency, and a manual-promote gate — turn a demo into something you can actually run. WhatsApp is documented as the upgrade path. (Scorecard: Cost-bounded + Governed.)
The flagship framework: five gates — Secure, Provable, Observable, Cost-bounded, Governed — that every AI system clears before I call it done. This is the writing-and-teaching spine that grades the agent from Lab #1 and connects every lab into a series. (Scorecard: all five.)
Failure as the lesson: watch a mediocre agent fail on camera, diagnose why, and convert those failures into graders. This instantiates the Provable dimension — turning vibes into evals that catch regressions before users do. (Scorecard: Provable.)
One task, built three ways — n8n, a raw Claude loop, and LangGraph — scored on one private eval harness with a public leaderboard. The heavy, maximally shareable anchor that makes the trade-offs between stacks provable instead of tribal. (Scorecard: Provable, with Cost + Governed.)