←Back to Projects
Flux: A Privacy-First, AI-Powered Language Learning Assistant
January 21, 2026

Flux: A Privacy-First, AI-Powered Language Learning Assistant

Bridging the gap between language learning and fluent reading, Flux is a "Self-Hosted Cloud" application. It leverages local LLMs (Ollama) and neural TTS to provide context-aware translations and an interactive audio experience, all while keeping data 100% private.

ReactNestJsOllama (Local AI)PostgreSQLNextJSPrisma View Code
Flux: A Privacy-First, AI-Powered Language Learning Assistant

Flux: Engineering a Private, AI-Driven Reading Experience

In an era of cloud-dependency and data tracking, Flux was born from a simple question: Can we build a premium AI reading assistant that runs entirely on local hardware? Flux is a sophisticated "Self-Hosted Cloud" platform designed to help language learners move from struggling with text to fluent comprehension through interactive AI tokenization and neural audio.

🧠 The "Self-Hosted Cloud" Architecture

Most AI apps are thin wrappers around expensive APIs. Flux takes a different approach by treating your local machine as the cloud provider.

  • Hybrid Monorepo: Using NPM Workspaces, I managed a React 19/NextJS frontend and a NestJS backend in a single repository.
  • The Backend "Brain": The NestJS server acts as a secure proxy. While the frontend feels like a native app, the backend handles the heavy lifting—managing persistence via Prisma and orchestrating requests to Ollama (the local LLM engine).
  • AI Abstraction: The frontend never talks to the AI directly. By proxying through NestJS, the system is future-proof; I can swap local models for OpenAI or Anthropic without changing a single line of frontend code.

✨ Engineering the User Experience

Building a tool for "Deep Reading" required solving complex UI and synchronization challenges that go beyond standard web development.

1. Interactive Tokenization

Text isn't just a string in Flux; it’s an interactive map. I developed a tokenization engine that breaks down text into selectable units. This allows for:

  • Context-Aware Dictionary: Hovering over a word doesn't just give a definition; it sends the surrounding sentence to the local LLM to provide a translation that fits the specific context.
  • Streaming AI Content: Users can generate custom reading material, which is delivered via NDJSON streaming for real-time feedback.


2. Neural Audio & Karaoke Sync

To aid pronunciation, Flux features a high-fidelity audio engine.

  • Karaoke Highlighting: Using the browser’s Speech Synthesis API, I implemented real-time word highlighting that stays perfectly synced with the audio playback.
  • Smart Resume: The system tracks playback position at the sentence level. If you pause, Flux intelligently resumes from the start of the last logical thought, ensuring natural continuity for the learner.

🛡️ Privacy as a Core Feature

The primary value proposition of Flux is Zero Data Leakage.

By containerizing the PostgreSQL database and leveraging local LLMs, Flux ensures that a user's reading habits, saved vocabulary, and generated texts never leave their network. It provides the "smart" features of a modern SaaS without the subscription fees or the privacy trade-offs.

🛠️ The Tech Stack

  • Frontend: React 19 (Signals-ready), TypeScript, Tailwind CSS, Zustand (State Management), Framer Motion.
  • Backend: NestJS, Prisma ORM, Express.
  • Infrastructure: Docker (Postgres), Ollama (Local LLM).

💡 The Takeaway

Flux demonstrates my ability to build vertically integrated products. It shows proficiency in modern React patterns, backend proxying, and the emerging field of local AI integration. It’s a project that balances complex technical requirements with a clean, user-centric design.

Crafted with ❤️ by Stanley Morales