OpenReader

A local-first AI reader that turns text, PDFs, EPUBs, and articles into synchronized audio on macOS and the web.

React
TypeScript
Tauri
Rust
FastAPI
MLX

2: Desktop + web apps
Local: Private AI inference
Solo: Built end to end

Challenge

Turn structurally different text, PDF, EPUB, and web content into one responsive read-along experience while running speech generation privately on the reader's device.

Responsibility

Sole designer and engineer. Built the shared reader engine, web and desktop applications, local inference backend, storage integrations, testing systems, packaging, and release pipeline.

Solution

Built a Bun monorepo around a shared React reader package, with Next.js for the web, Tauri and Rust for macOS, and a FastAPI backend using MLX and PyTorch to run Kokoro and Qwen3 locally with direct word timestamps.

Outcome

A live local-first product built independently from end to end, with private offline generation, synchronized word highlighting, project persistence, media export, and shared experiences across macOS and the browser.

Engineering decisionsD-01

One reader engine, two applications

A shared React package powers both the Next.js web application and the Tauri desktop application, keeping reading, generation, playback, and project behavior consistent across surfaces.

Generate in prioritized chunks

Audio generation follows reading priority so playback can begin before an entire book is processed. Cancellation, debouncing, precomputed word mappings, and dual audio elements keep seeking and playback responsive.

Package the runtime, download the models

The desktop application bundles its inference runtime while model weights are installed separately, preserving private local execution without making the distributed application unnecessarily large.