Hey dev.to — I'm Lepp, and I've spent the last few months building Neuro-Kindred, a free web app (installable PWA) for neurodivergent adults (18+) to find friends or dates. I built it solo on an AI app-builder called cto.new, which means my whole workflow looks a little unusual: I write the specs, the builder's AI writes the code, and then I make it try to break its own work in a deep debug audit.
I'm sharing the interesting product and technical choices, because a few of them were genuinely hard to get right.
The app
https://neuro-kindred.com/discover — and a no-account-needed demo with sample profiles and playable games: https://neuro-kindred.com/demo
Feature list, briefly: context-rich profiles, an AI bio helper and AI match explanations, messaging, voice/video calls, Connect Four and group trivia games, watch-together rooms, calm accessible design. Public pages in six languages (English, Español, Français, Deutsch, Português, हिन्दी). Sign-up is username/email/password/date-of-birth only — no country field, open worldwide. No paywall, no subscriptions; there's an optional one-time tip jar ($2/$5/$10/$20) that nobody has used yet. Honest status: 13 real members, $0 revenue.
The interesting choices
1. "Shiny pebbles" instead of likes. No like counters, no match percentages, no read receipts pressure. You give someone a pebble — a small, deliberate signal. The reason it's a technical story: double-sending a pebble was a real race condition (two taps, two pebbles, awkward). The fix went to the database layer — partial unique indexes plus a Postgres trigger guarding the count. Same treatment for game room joins and room caps. It taught me that for a social app, every interaction counter is a race condition waiting to happen, and the only place to fix it for real is the DB.
2. Read-only demo mode. The demo is a fully interactive tour with sample profiles, games, and a watch room — but every locked action opens a "create a free account" dialog instead of writing to the database. The audit verified zero DB writes from the demo. It's a meaningful amount of engineering to let strangers click around a social app without creating accounts, and it was worth it: it's the first thing I'd send anyone.
3. A three-strike moderation system, by design not by ML. I'm building this right now: strike one is an automatic in-app warning, strike two is a final warning, strike three (from a third different member) deletes the account. The design constraints were the interesting part — a strike counts only from a distinct reporter, so one person can't mass-report someone out of existence; self-reports don't count; admins get a pardon/reset override; warning copy had to be written in all six languages with the same calm tone. No email system exists, so everything is in-app. Coordinated-report abuse is the residual risk and I'm not pretending it's solved.
4. Language support that doesn't break the PWA. The picker sits on all eight public pages, works logged-out, honors ?lang= parameters, and persists via a cookie. Partial translation is worse than none if it surprises people, so I documented exactly what stays English (the demo is effectively English-only for now) rather than claiming full coverage.
Lessons
- The audit habit compounds. The deep debug I run on this app (build, 200+ tests, secrets scan, route checks, protected-API checks, DB census) catches the boring defects that would be catastrophic in a social app: orphaned rooms, race conditions, missing report paths. Spec the audit, don't just ship.
- AI builders move fast and break social invariants. The fastest code to write is the most optimistic code — no double-click handling, no concurrent-join logic. For a social app, "what happens when two people do this at once" is the first question, not the last.
- Distribution is the actual hard part. The build was the fun half. Thirteen real members came one at a time through directories, permission-first subreddit posts, and international ND communities. No trick, just slow.
If you poke at the demo, I'd genuinely love feedback on two things: (1) what's confusing in the first 60 seconds, and (2) any social-invariant edge cases I clearly haven't thought of — you all are good at that one.
— Lepp

