<aside> ⚠️

Date: Sunday, October 25, 2026 · Week 4, Day 28 · Type: Build Day + Live Session 8 (Capstone)

Scheduling note: the landing page copy labels Live Session 8 as falling on a "Friday." Counting 30 straight days from the confirmed Day 1 (Monday, Sep 28) puts Day 28 on Sunday, October 25, not a Friday — see the flagged issue on the main roadmap page. Confirm the actual calendar invite time before this goes out to the cohort.

</aside>

Learning Manual

28.1 The AI System Design Interview — how senior AI engineering candidates are evaluated at OpenAI, Anthropic, Google, and Palantir.

28.2 The 4-Step System Design Framework — Requirements & Scope (QPS, Latency, Cost) → High-Level Architecture & Flow → Deep Dive into Core Components (Vector DB, Caching, LLM Serving) → Bottlenecks, Failure Modes & Mitigations.

28.3 Paper Deep Dive — Efficient Memory Management for Large Language Model Serving with PagedAttention (Kwon et al., 2023).

The Daily Build

Create a complete System Architecture Blueprint (using Excalidraw or text-based diagrams) and a 2-page technical design document for an enterprise AI document assistant serving 10,000 corporate users.

Architecture Diagram

+-------------------------------------------------------------------------------------------------------+
|                              ENTERPRISE 10,000-USER ARCHITECTURE BLUEPRINT                            |
|                                                                                                       |
|  [10,000 Users] ---> [Load Balancer] ---> [FastAPI Gateway Cluster]                                  |
|                                                  |                                                    |
|                  +-------------------------------+----------------------------------+                 |
|                  |                               |                                  |                 |
|                  v                               v                                  v                 |
|       [Redis Semantic Cache]            [Guardrail Security]             [Hybrid Search Vector DB]    |
|                  |                               |                                  |                 |
|                  +-------------------------------+----------------------------------+                 |
|                                                  |                                                    |
|                                                  v                                                    |
|                                    [LLM Inference Cluster / APIs]                                     |
+-------------------------------------------------------------------------------------------------------+

Interview Questions

  1. Walk me through your system design for an AI customer support bot handling 10,000 concurrent requests per minute. How do you prevent API rate limit crashes?
  2. Where would you place a caching layer in an enterprise RAG architecture, and what is your cache invalidation strategy?
  3. In the PagedAttention paper, how does virtual memory pagination prevent GPU memory fragmentation during generation?

Problem Statement

Problem 28. Write a Python capacity estimation script estimate_cluster_capacity(total_users: int, queries_per_day: int, avg_tokens: int) that calculates required queries-per-second (QPS), estimated daily token consumption, and monthly hosting costs.

AI-DSA Primitive

Data Structure: Consistent Hashing Ring.

Mechanics: Circular hash ring mapping keys and server nodes to balance load and minimize re-mapping during server scaling.

Real-World Production Use: Deployed inside distributed vector databases (Qdrant, Milvus, Pinecone) to shard vector collections across server clusters.

Software Design Pattern