UNJYNX
METAminds — AI Productivity Platform
The Challenge
Traditional TODO apps all make the same fatal assumption: that users will open them. They will not. People live inside WhatsApp, Telegram, Instagram, and Slack all day — bouncing between conversations — and productivity apps insist on being one more tab they have to remember to check. Reminders get buried in notification trays, swiped away, or simply ignored. Meanwhile, existing tools offer no escalation when tasks are missed, no native support for the 8 channels where people actually communicate, and no AI that can parse messy natural-language task entry. The satisfactory baseline for productivity software has not moved in a decade.
The Approach
Built UNJYNX as a complete productivity operating system spanning 3 independent repositories and 6 containerized services on a Hetzner VPS. Engineered a Flutter monorepo organized into 24 Melos packages with a plugin-play architecture — so every feature is a self-contained package with its own ports, models, and screens. Backed by a Hono + TypeScript API with 300+ endpoints across 37 modules and 15 middleware layers. Paired it with a Next.js 15 + React 19 web client and an Astro landing page on Vercel. Wired in a self-hosted Logto OIDC server for authentication (Google OAuth + MFA), Drift SQLite offline-first storage with field-level LWW conflict resolution, and an internal ML service running Thompson Sampling + LinUCB bandit algorithms for personalization.
Multi-Repository Architecture
Six Docker containers on a single Hetzner VPS: backend API, Logto auth, web client, PostgreSQL, Valkey cache, and an internal ML service. The Flutter mobile app is built from a 24-package Melos monorepo. Landing page runs separately on Vercel.
Flutter Monorepo (24 Packages)
Cross-platform Flutter app (Android + iOS) organized as 24 Melos workspace packages with plugin-play architecture: core (theme, models, 7 port contracts), 18 feature packages, and 5 service packages. Drift SQLite offline-first with 12 local tables and field-level LWW (Last-Writer-Wins) conflict resolution. FCM + awesome_notifications for push. Dio HTTP client with Logto OIDC auth. 61+ screens.
Next.js 15 + Astro
User web client built on Next.js 15 + React 19 (38 pages covering tasks, projects, AI, channels, billing, admin). Astro landing page on Vercel for marketing at unjynx.me. Admin portal (admin.unjynx.me) and dev portal (dev.unjynx.me) built as React SPAs deployed through Nginx on the VPS.
Hono API + Logto Auth
Hono + TypeScript backend exposing 300+ REST endpoints across 37 modules with 15 middleware layers. Self-hosted Logto OIDC server (Google OAuth + email/password + MFA) on port 3001-3002. Neon (managed Postgres) for primary data, local PostgreSQL for content, Valkey for Redis-compatible caching. Internal ML service runs Thompson Sampling + LinUCB bandit algorithms for reminder timing and channel selection.
The Solution
- Built 24 Melos packages: core, 18 feature packages, and 5 service packages (api, auth, database, sync, notification)
- Engineered a plugin-play architecture with 7 port contracts (Auth, DB, Sync, Notification, Storage, AI, Preferences)
- Delivered 8 notification channels: WhatsApp, Telegram, Instagram, SMS, Discord, Slack, Email, Push
- Implemented smart escalation chains: task → gentle ping → firm reminder → accountability partner → public shaming (opt-in)
- Integrated 5 AI personas (coach, friend, drill sergeant, zen, chaotic) for personalized chat and task reasoning
- Built comprehensive gamification: XP, achievements, leaderboards, challenges, accountability partners, activity heatmap
- Designed offline-first sync with Drift SQLite + push/pull sync engine + LWW conflict resolution + exponential backoff
- Shipped 6 containerized services on Hetzner: backend API, Logto, web client, Postgres, Valkey, ML service
- Built 4 separate portals: user (app.unjynx.me), admin (admin.unjynx.me), dev (dev.unjynx.me), landing (unjynx.me)
- Included sprint boards, burndown charts, velocity tracking, and retrospectives for team workflows
- Added hierarchical goal tree, time blocking, pomodoro, morning/evening rituals, ghost mode, weekly review
- Implemented ML-driven reminder timing via Thompson Sampling + LinUCB contextual bandits
Feature Breakdown
Core Productivity
- Task CRUD with Kanban / timeline / table views
- NLP task parser (dates, priorities, channels from natural language)
- Hierarchical goal tree with progress tracking
- Sprint boards with burndown, velocity, retrospectives
- Time blocking + pomodoro + morning/evening rituals
- Recurring task builder with timezone handling
8-Channel Notifications
- WhatsApp, Telegram, Instagram, SMS, Discord, Slack, Email, Push
- Smart escalation chains (soft → firm → accountability)
- Channel preference routing (work → Slack, personal → WhatsApp)
- Quiet hours with timezone awareness
- Test send + delivery history
- FCM + awesome_notifications + deep link tap handler
AI & ML
- 5 AI personas (coach, friend, drill sergeant, zen, chaotic)
- Thompson Sampling bandit for reminder timing
- LinUCB contextual bandit for channel selection
- AI insights and weekly reviews
- Team AI for shared workspace intelligence
- Industry-specific modes with tuned defaults
Gamification & Team
- XP system + achievements + leaderboards
- Challenges + accountability partners
- Activity heatmap + progress charts
- Team dashboard + shared projects + standups
- Team messaging + DMs + shared channels
- Org onboarding + invite accept flow
Scale & Scope
The Impact
Tech Stack
“The best TODO app is the one that finds you where you already are. UNJYNX does not ask users to open an app — it reaches them across WhatsApp, Telegram, Instagram, SMS, Discord, Slack, Email, and Push, through smart escalation chains personalized by Thompson Sampling and LinUCB bandit algorithms. Built as 3 repositories and 6 Docker services on a single Hetzner VPS, with 24 Flutter packages, 300+ endpoints, 37 backend modules, and a self-hosted Logto OIDC auth server, UNJYNX is what productivity software looks like when you refuse to accept that the satisfactory baseline is acceptable.”