Private matrimonial platform

Consumer

HeyLayla

A privacy-first matrimonial platform where consent and disclosure are enforced by the application, with a controlled conversational companion and an operator research capability built on hybrid retrieval.

Status
Live, pilot cohort
Retrieval
FTS + pgvector HNSW + exact reference, RRF-fused
Prompt law
Three inputs only, sentinel-tested
Decisions
60+ architecture decision records
Stack
TypeScript, React, Hono, Drizzle + Supabase Postgres, pgvector, pg-boss, Fly.io, Vercel, Astro, Playwright

The problem

Younger adults in socially conscious markets want to choose their own partner seriously, calmly and privately. Incumbent matrimonial sites sell trust to families through community data and paywalled contact details; dating apps sell choice through swipe mechanics that quietly corrode serious intent. Neither answers the person who wants to run their own search with dignity.

What we built

  • A TypeScript product monorepo: a member app and admin console in React, an API in Hono with cookie sessions, Supabase Postgres through Drizzle with a deny-all row-level-security backstop, a queue and single worker for the photo pipeline and transactional email, and an Astro marketing site.
  • A staged self-disclosure engine at the core of the product. One function decides what a viewer may see of a subject; nothing else is allowed to.
  • Scarce, mutual introductions modeled as a state machine, automatic verification on data completeness, kind closure as the default, and a permanent record of relationship milestones.
  • Transactional email enqueued in the same database transaction as the business event, drained by the worker with leases, backoff and dead-lettering, and reconciled from delivery webhooks.

Where AI fits

  • Layla, a conversational companion, appears only when invited. Her prompt is assembled from exactly three inputs: the member's own words, the counterpart fields the disclosure engine already granted, and the conversation both members share. Anything she proposes lands in the member's own draft; she never sends.
  • Engineering Memory, an operator-only research surface over the project's issue history and repository documents: asynchronous idempotent ingestion, deterministic chunking, Postgres full-text search combined with pgvector semantic search and exact-reference lookup, fused with weighted reciprocal rank fusion, cited answers streamed over server-sent events, and lexical-only operation when embeddings are unavailable.

Engineering discipline

  • Fail closed everywhere: malformed or out-of-enum data resolves to least visibility or throws, never falls through open.
  • Architecture decisions are recorded before implementation. The repository carries more than sixty ADRs, and its positioning document is treated as a constitution every product decision traces back to.
  • Dependency boundaries enforced by tooling, red-first tests, and a two-member end-to-end journey that replays the whole product arc from sign-up to a kept story.

What it is worth

  • Members get verification as the admission floor, private discovery and staged disclosure without a family committee or a feed.
  • Operators can answer 'what did we decide, and why' with citations instead of archaeology.

The two lanes

Kept deterministic

  • Disclosure and visibility
  • Consent and authorization
  • Introduction state machine
  • Verification and activation

Where intelligence helps

  • Conversation guidance, on request
  • Profile writing help
  • Operator research with citations