Four crossroads I encountered while building an AI paper tool

The author built a local-first AI tool for reading and translating English papers in a single view.

During development, they hit issues with two-column PDF extraction, backend freezes during embeddings, and 524 timeouts from OpenRouter/Gemini.

Rather than follow generic AI advice, they stabilized the tool with blank-area detection, one-piece-at-a-time processing, and streaming responses.

The project shows that turning a prototype into a real product requires balancing cost, reliability, and user experience.