The collection
Inside AI Systems — One Problem, One Architecture, One Lesson.
A daily educational reading series: investigate one practical problem, trace an architecture, and test what you learned. Start with seven complete lessons about an AI student assistant.
For: Students and engineers learning how AI systems work
Prerequisites: Basic familiarity with an AI assistant; no GPU required
Your first week inside an AI system.
Follow an AI student assistant from its first traffic spike to a controlled release. All 7 complete lessons are available now. “Day” is a suggested reading pace.
Day 01 · Program 03
Your AI Model Is Fast. Why Are Users Still Waiting? →
The problem: Responses slow down after a webinar brings a traffic spike.
The lesson: Measure waiting, processing, generation, and delivery separately.
Day 02 · Program 03
Why AI Requests Wait: Bounded Queues and Deadlines →
The problem: An unlimited waiting room turns a short burst into prolonged overload.
The lesson: Bound the work you accept and propagate the time you have left.
Day 03 · Program 01
When a Document Assistant Answers Without Evidence →
The problem: A confident answer cites an outdated or inaccessible policy.
The lesson: Retrieve authorized, versioned evidence and allow the system to abstain.
Day 04 · Program 01
Why an AI Agent Creates the Same Ticket Twice →
The problem: A tool call succeeds, its reply is lost, and a retry creates a duplicate.
The lesson: A timeout means an unknown outcome; make consequential actions recoverable.
Day 05 · Program 06
Why More AI Replicas Do Not Immediately Fix a Spike →
The problem: The autoscaler requests capacity while users are still waiting.
The lesson: Requested capacity becomes useful only after startup and readiness.
Day 06 · Program 06
When Caching Makes an AI Answer Fast and Wrong →
The problem: A cached answer survives a document update or permission change.
The lesson: A cache hit must satisfy the same scope and freshness contract as a fresh answer.
Day 07 · Program 08
How to Release an AI Change Without Guessing →
The problem: A faster prompt passes a few examples but breaks important user tasks.
The lesson: Release a versioned system against explicit quality and operational gates.