System Design & Case Studies

Choosing a vector database, an evaluation strategy, or an agent architecture is a system-design question, and the strongest way to prepare for that question is seeing how real companies actually answered it — what constraint drove the decision, what the alternative would have cost them, and what broke before they got it right. Every case study here is built from a real, verifiable primary source (an engineering blog, a peer-reviewed paper, an official incident report), not marketing copy — and where a claimed number couldn't be verified, it's flagged rather than presented as fact.

Retrieval infrastructure at scale

• Notion → turbopuffer — a cost- and multi-tenancy-driven migration at real SaaS scale, 600× growth in onboarding
• Uber's billion-scale vector search — what happens when an HNSW graph stops fitting in memory
• Airbnb: choosing IVF over HNSW — a real counter-example to reflexively picking the "better" algorithm
• Spotify: Annoy → Voyager — when deployment constraints, not benchmarks, decide the architecture

Production RAG and retrieval architecture

• GitHub Copilot's embedding model — why hard negatives, not easy ones, are the real retrieval bottleneck
• LinkedIn's GraphRAG for customer support — a peer-reviewed system with a rare real business-impact number
• Dropbox Dash: RAG to agents — an architecture that evolved in public, including a "more tools made it worse" finding
• Sourcegraph Cody: abandoning embeddings — when the theoretically-best retrieval method isn't the deployable one

Agentic systems and hybrid search in production

• Klarna's AI customer service agent — real numbers, and the honest walk-back a year later
• Uber QueryGPT: multi-agent text-to-SQL — decomposing one agentic task into narrow, evaluable sub-agents
• Perplexity's hybrid search at web scale — hard evidence that sparse and dense retrieval are complementary, not redundant
• Shopify's real-time product search — why hybrid search needs a streaming embedding pipeline, not a static index

Cost and evaluation in production

• Dust: tiered prompt caching — segmenting a cache by how often each layer actually changes
• Anthropic's own eval postmortem — three real regressions that shipped past Claude Code's own eval suite
• Notion's dual-track evaluation framework — separating "did we break anything" from "is the new model better"

Security and document processing at scale

• EchoLeak: a zero-click indirect injection — the full attack chain behind CVE-2025-32711
• Anthropic's Constitutional Classifiers — a jailbreak defense that got 40× cheaper without getting weaker
• Harvey AI: document processing at legal scale — what breaks when ingestion volume grows 26× in a year