Beaver AI

Payment webhook idempotency via Redis

Every inbound event ID cached for 24h before processing. Replays return early without side effects. The kind of thing you only learn to do after a duplicate charge.

Subscription cache sentinel pattern

Store sub or False, return cached or None. Distinguishes "we've checked and there's no subscription" from "we've never checked" without an extra DB hit on every authenticated request.

Custom RAG with per-user vector indices

client_document_id = meeting_{pk}_user_{user_pk} ensures clean per-user semantic search with deterministic dedup keyed on the meeting PK. No cross-user bleed; safe re-ingest.

Cascade-protect on the user FK from meetings

Explicit choice - prevents a cascade from accidentally wiping a user's entire meeting history. Forces explicit handling of user deletion, which is the right scaffolding for a GDPR-aware product.

Compound indexes matched to actual query patterns

(user, -started_at) on Meeting and (meeting, start) on TranscriptSegment. Replaced three Python-side filtering patterns with three indexed queries when the data outgrew memory.

Drip email idempotency

A drip-log table with a (user, day_number) unique constraint stops the hourly task from sending duplicates if it overlaps itself.

Two distinct OAuth flows, one app

The signup OAuth library handles auth identity. The calendar integration is separate - different scopes (calendar.readonly), separate refresh-token storage on its own integration record. Conflating them is a footgun; separating them is correct.

Multiple auth backends, explicit

When both a password backend and a third-party OAuth library are active, every login call has to name the backend explicitly. Easy omission, baffling failures.