Velvet AI Companion

Velvet is an AI companion platform where users are matched to one of 8 distinct archetypes based on preference, gender, and identity inputs. Each archetype has a specific behavioral profile, STAGE_INSTRUCTIONS for personality progression, and conversation context injected into every OpenAI call.

The platform is live on web and submitted to Google Play as a Capacitor-wrapped Android app. It has a full test suite, session isolation for concurrent users, and a progression engine that deepens the companion's behavior at 8, 20, and 40 message thresholds.

• 8-Archetype Matching Engine

  Users complete a matching flow that collects preferences, identity inputs, and communication style signals. The engine routes each user to one of 8 archetypes and loads the appropriate STAGE_INSTRUCTIONS, companion name, and behavioral context. Preferences include identity and gender expansion options.

• Session Isolation

  Every user session is scoped to an httpOnly cookie. Conversation history, session state, and resets are all per-user — no shared global state. 18/18 isolation tests confirm that resetting one user's session has zero effect on any other active session.

• Behavioral Progression System

  The companion's personality depth increases at 3 thresholds: 8, 20, and 40 messages. STAGE_INSTRUCTIONS are injected into the system prompt at each threshold, shifting tone from warm and exploratory to more personal and emotionally attuned. 78 behavior tests cover the full progression arc.

• Android Build via Capacitor

  The Next.js web app is wrapped as a native Android app using Capacitor. The AAB build (2.85MB) was submitted to Google Play closed testing. A VBS wrapper was written to run scheduled background tasks without visible terminal windows.

• Session reset wiping all users

  The reset endpoint was clearing a global state object instead of scoping to the requesting user's session token. During concurrent QA testing, one tester's reset wiped every active session. Fix: scoped reset to the httpOnly cookie — fully isolated per user.

• 22 quality tests for conversation safety

  The OpenAI responses needed to stay within character and not break companion voice under edge cases. Built a 22-test quality suite covering response boundaries, character consistency, and stage transition accuracy.

• Back button exiting the app on Android

  The Android back button was sending users out of the app entirely instead of navigating within the companion flow. Documented as BUG-01 before Play Store submission. Session trap fix was built and committed (not yet deployed at time of submission) to be resolved before public release.

Live at emotional-companion-pi.vercel.app. Android AAB in Google Play closed testing. 78/78 behavior tests passing, 22/22 quality tests passing. Session isolation verified.