Serenia

Problem

Most AI assistants are optimised for usefulness: precise, formal, and slightly exhausting. Serenia started from a different question — what if the goal was to feel like talking to a close friend?

The app gives the AI a distinct personality: casual, occasionally sarcastic, never verbose. Messages are capped at 180 characters. It’s closer to an SMS thread than a knowledge base.

Design decisions

Tone over features. The product surface is deliberately small — chat, history, a subscription. The investment went into making the AI’s character feel consistent and human. That meant a lot of prompt engineering and a lot of reading responses that felt slightly off.

Security at the base layer. Conversations are sensitive. The architecture uses user-derived encryption keys (HKDF) so that message history is encrypted in a way the server cannot read. JWT authentication over RSA, TLS 1.3 throughout.

Mental health guardrails. When the model detects distress signals in a conversation, it automatically surfaces the national suicide prevention line (3114). This was non-negotiable from the start — a “friend” app has a duty of care that a general assistant doesn’t.

Stack

• Backend: Java 21 + Quarkus 3, Hibernate Panache, Liquibase migrations
• Frontend: Angular 21, TailwindCSS 4
• Database: PostgreSQL 16
• Infrastructure: Docker, Traefik, Nginx
• External: OpenAI API, Stripe for subscriptions

Status

Live at serenia.studio. Two stars on GitHub, four open issues, and a growing list of things I want to change now that real users are using it.