Village Hale — the AI village for families

Hale isn’t a chatbot you open. It’s a passive system: the web app enqueues events, a worker drains them off a Postgres-backed queue, and a pipeline of specialised agents classify, infer, draft, and review — each testable on its own, none trusted on prose alone.

The LLM is maybe a fifth of the system. The rest is orchestration, guarded tool interfaces, audit logging, and evals — the parts that decide whether an agentic product is trustworthy or a demo.

• web ↔ worker split across one typed event contract
• the reviewer agent must invoke verification tools before it can approve
• every action writes an immutable audit row

This handles newborn and child data, so the default is the most restrictive thing that still works. Discovery sees only a coarse area— a postal prefix, never an address. A teen’s raw content is redacted from parents by default; only a category surfaces, and raw access is a logged, time-limited grant.

It’s enforced in code, not on a privacy page: a render-strip test fails the build if a child’s name could reach a session recording, and redaction sits at context-assembly so it can’t be forgotten downstream. PIPEDA, Quebec Law 25, Canadian data residency — the compliance core, not the fine print.

Every other product leads with the assistant. Hale leads with the village— the genuinely good local things families near you actually do, vouched for by parents like you, more useful with every family that joins. The concierge finds and organises it; the network is what compounds.

Recommendations are grounded and cited — real venues, real sources — never a confident hallucination about where to take a toddler.

New families start at observe-only. Hale drafts; the parent approves every action. Only after five clean approvals of one kind of task may it handle that kind on its own — and any automation can be withdrawn instantly. It never spends past a cap, or acts across both parents’ data without consent.

A household assistant that acts without permission isn’t a feature — it’s a liability. Trust is the product.

Every agent role has an eval — real cached responses, no mocked LLMs — that scores it across models. When “just use the biggest model” was tempting, the matrix showed Sonnet matched Opus on every role, and that Haiku silently failed the teen-safety classifier. So the tiering is measured: the cheapest model that holds the bar, the strongest only where it provably helps.