Part 16: Case Studies
Eight production AI architectures, reconstructed from public engineering blogs, conference talks, and podcast interviews.
In one line: The earlier chapters tell you what primitives exist; this chapter shows you how shipped products in 2026 combine those primitives into specific architectures — what's durable, what's a snapshot, and what you'd copy vs. avoid.
Reading about primitives in isolation tells you what's possible; reading how real products combine them tells you what's practical. This chapter is eight worked architectures of shipped AI products — four developer-tooling, two enterprise-vertical, one productivity, one consumer. The shape of each (high-level architecture, key engineering decisions) is durable across model changes; the specific tools each uses (which model, which vector DB) are 2026 snapshots and will rot. Read the shapes, not the brand names.
How to read this chapter
Each case study follows the same structure:
- Product in one paragraph. What it does, who it's for.
- Architecture diagram. The durable shape.
- Two or three key engineering decisions. What's non-obvious and why.
- Stack snapshot (2026). The specific tools they're known to use as of this writing. Will age.
- What to copy / avoid. Practical lessons.
- Sources. Public talks, blogs, interviews.
Read them in any order. They're independent.
The eight case studies
Developer tools
- Cursor — IDE-native coding assistant. Context retrieval, apply-diff, predictive autocomplete.
- Claude Code — Terminal-native coding agent. MCP-first, tool discipline, agent loop.
Search & answer engines
- Perplexity — AI answer engine. Real-time web retrieval, citation enforcement, multi-step research.
Voice & customer support
- Sierra — Customer-service agents. Published task and supervisor architecture, voice latency, and a traced return-request example.
Vertical AI (enterprise)
- Harvey — Legal AI. RAG over case law and contracts, citation-grade retrieval.
- Glean — Enterprise unified search. Per-tenant + per-user ACLs, federated retrieval.
Embedded AI in productivity
- Notion AI — AI inside a productivity app. Per-block actions, document-aware context, no chat-first paradigm.
Consumer AI
- Duolingo Max — AI-powered language tutoring. Cost-controlled per-turn, persona-driven, eval-heavy.
What's durable vs. what's a snapshot
The pattern across all eight: there's a durable architecture that has been roughly the same shape for ≥18 months despite model upgrades, and a tools snapshot that changes per quarter.
| Durable (years) | Snapshot (months) |
|---|---|
| The retrieval shape (chunking strategy, hybrid, reranker) | Which embedding model |
| The context-management strategy | Which model version |
| The eval pipeline | Which observability platform |
| The auth & permissioning model | Which vector DB |
| The fallback ladder | Which gateway / proxy |
| The latency budget | Which inference provider |
When you read these case studies, the architecture diagrams are the load-bearing knowledge. The stack snapshots are the disposable layer.
What these case studies are NOT
- Endorsed by the companies. All reconstructed from public sources — engineering blogs, conference talks (AI Engineer Summit, Latent Space), podcasts, founder interviews, GitHub commits where the code is public.
- Complete. Every shipped product has internal complexity these summaries skip. You're getting the architecture shape, not the implementation manual.
- Static. All of these companies ship continuously. Specifics will shift; the shapes should outlast the shifts.
What you should get out of this chapter
- Pattern fluency. Recognize architectural choices in the wild and know what they cost / what they win.
- Decision vocabulary. "Like Cursor's apply-diff but with our codebase indexer" beats reinventing terminology.
- Interview material. AI system-design rounds (see AI system-design interviews) often anchor on a real product. Knowing one of these case studies cold lets you say "I'd model this on how Glean handles ACLs, with these adjustments…"
- Inspiration with calibration. Most of these took 50–500 engineers to build. Knowing what's possible is good; knowing what's practical at your headcount is better.
→ Start with Cursor.
What's next
When you've worked through the case studies that interest you, the optional Cutting Edge chapter (Chapter 17) maps emerging themes — agent harnesses, agentic RAG, trajectory evals — without requiring them to ship today. Then take the Final capstone or keep the Glossary open as your standing reference.