The consumer assistants people use every day are the largest-scale agentic systems running anywhere. This knowledge base explains how they are built — the per-turn context pack, memory namespaces, the stateless fleet and its control plane, token economics, quality gates and trust — and then maps each piece one-to-one onto enterprise systems: tenants, policies, audit, residency, human-in-the-loop. Same chassis, stricter constraints. Every concept was proved on a working reference build first.
Read it left to right: a person reaches an edge; an identical, stateless worker assembles that person's context pack from shared stores, asks a routed model service, optionally calls tools, and replies. Whatever the person revealed is written back to memory after the reply. Around it, a control plane sizes the fleet with one small loop — and every token is accounted for.
Every consumer-AI team hits the same wall: one assistant has to feel personal to each of millions of people, and the bill is model tokens, not servers. Most public material explains either the model or the product; almost none explains the architecture in between — where personalization actually lives, what a worker holds and what it must not, how a fleet is sized, and why “scale” is a reconcile loop rather than a bigger box.
This site is that missing layer, written the way a principal engineer would explain it to a smart colleague from another discipline — one idea per section, a diagram wherever a diagram beats a paragraph.
Personal is a per-turn context pack: exact profile facts by keyed lookup, a tone template chosen by segment, and top-K memory retrieved inside the person's own namespace.
Scale is identical stateless workers plus a control plane whose reconcile loop keeps the fleet at desired = ceil(active / users-per-instance) — the Kubernetes HPA loop, made legible.
A user namespace becomes a tenant; a segment template becomes a policy; the same eval gate, observability and autoscaler apply — with residency, audit, SSO and human-in-the-loop layered on.
Guardrail → session → rewrite → retrieval → context pack → tools → generate → write-back, as a sequence diagram with latency and cost per stage.
Namespaces as the privacy boundary, write-back off the hot path, the 1 → 1B ladder, and the reconcile loop that scales out and in.
Routing, tiering, prefix caching, fallback and canary; a cost-per-turn table at 1K / 1M / 100M daily users.
Golden sets with an empathy score, SLOs and error budgets, guardrails in and out — then two worked enterprise scenarios (a bank's advisor, a retailer's support agent).
Architects and hiring managers in consumer or enterprise AI who want to see how someone reasons about scaling personalized AI; engineers moving from web-scale services to AI (it is the microservices discipline with five things promoted to first class); founders who need to know why personalization is a data-architecture problem.