AI School 369 · UrbanSchool369.com
Production AI, explained simply.
Generative AI, agentic AI, RAG, and evals, taught the way production teaches them: the failures, the fixes, and the working code. One concept, one working project, every edition. Every claim on this site traces to a runnable project or a real run.
Join Layer by Layer free: a new edition every Tuesday and Friday, 8 AM Central. Free forever. Unsubscribe anytime.
Pick your path
Preparing for AI interviewsScenario-first: the tool loop, graceful degradation, confirmation gates, and the free 100-question pack.Shipping a RAG systemFreshness, chunking, reranking, retrieval quality, in the order production teaches them to you.Building agents that behaveLoops you can bound: budgets, retries, circuit breakers, memory you can inspect.Keeping AI honestCanary evals and monitoring that tell you the system broke before your users do.
From the school
All articles →Agent EngineeringThe Agent Tool Loop, Taken ApartEvery agent is a loop: think, call a tool, read the result, repeat. The loop is simple. Keeping it from spinning forever is the engineering. Budgets, retries, and a circuit breaker, in a runnable demo.2026-10-23 · 7 min readRAG SystemsReranking: The Second Pass That Finds RelevanceYour retriever pulled 50 chunks and the right answer ranked 37th. Retrieval finds candidates; reranking finds relevance. Two-stage retrieval, explained with the scores on the table.2026-10-23 · 6 min readAI Trends & NewsWhat changed in AI this week, from primary sourcesLab blogs, arXiv, and Hugging Face papers, swept daily and published only after a human reads them. No rewritten press releases.Read the briefing →
The five workshops
Retrieval that stays honestRAG SystemsFreshness, chunking, reranking, retrieval quality. The failure modes nobody demos and every production system has.Autonomy, earned per actionAgent EngineeringTool loops, budgets, circuit breakers, confirmation gates, memory you can inspect. Agents you can bound are agents you can trust.Know before your users doEvals & ReliabilityCanary evals, LLM-as-judge failure modes, and the monitoring that tells you the system broke first.Architecture with receiptsAI System DesignThe design interviews test and production demands: latency budgets, cost routing, failure design, scale.The layer everything stands onData Engineering for AIPipelines, embeddings at scale, data quality. AI systems are data systems with a model attached.Watch insteadA Short every morningOne production AI concept in about a minute, daily at 7 AM Central, plus interview scenarios every Thursday and Sunday.
Projects you can run
All projects →RAG Freshness: Catch a Stale Index LyingA 260-line demo that builds an index, edits a document behind its back, and watches the chatbot answer from the old version. Then the fix: version stamps, hash-based refresh, deletion propagation, and 20 canary questions.Receipt: Verified run: 20/20 canaries pass on the fresh index, the stale index answers '30 days' after the policy changed to 60, refresh fixes it, deleted documents stop answering.Agent Tool Loop: Budgets, Retries, a Circuit BreakerThe think-act-observe loop in about 220 lines of standard-library Python, with the three controls production loops need: a step budget, a per-tool retry budget, and a circuit breaker.Receipt: Verified run: healthy task completes in 4 calls with a flaky tool succeeding on attempt 3; the stuck task's breaker trips at step 3 of 12.
Latest on the channel
All videos →Agentic AI Interview ScenariosInterview Scenario #1: The Agent Tool LoopYour agent has called the same tool 40 times. What went wrong, and what do you check first?Agentic AI Interview ScenariosInterview Scenario #2: One Tool Down Mid-TaskGraceful degradation: what a senior answer sounds like when a dependency dies mid-task.Agentic AI Interview ScenariosInterview Scenario #3: The Wrong-Flight BookingConfirmation gates and blast radius. Autonomy is earned per action, not granted per agent.