Decoded from public signals.
Each teardown analyses a public architecture from public signals only — engineering blog posts, conference talks, job ads, GitHub, observable behaviour. Diagrammed, decoded, with what to steal and what to avoid. No NDAs. No private conversations.
Claude.ai — the streaming-first AI product.
Plane separation (app vs AI), async safety telemetry, model gateway, MCP runtime, per-decision audit. The pattern Anthropic ships in its own product is the right substrate for enterprise GenAI in 2026.
Linear — the local-first sync engine.
Custom sync engine over generic CRDT. WebSocket transport. Workspace logical sharding with cell-readiness. Two protocol surfaces (GraphQL for partners; bespoke for client). The substrate that makes Linear feel instant.
Notion — one decision, ten years of consequences.
Block model drives every product superpower and every engineering pain. Per-block ACLs. Postgres re-shard journey (1 → 32 → 96 → Citus). Notion AI as RAG over existing blocks. Canonical block-store reference.
Vercel — the edge platform that earned its primitive.
Immutable atomic addressable deploys are the primitive everything else derives from. Fluid Compute consolidated the runtime. Routing Middleware generalised. AI Gateway as platform tier. Lessons for any internal platform team.
One teardown per quarter. On the queue: Stripe Billing · Cloudflare R2 · Supabase · Anthropic API gateway internals (from MCP signals). If there's one you'd want first, tell me.