Babilo DevLog #3: Asynchronous Engines and UX Refinement

While we have been defining the Learning Modes specification, the core engine has undergone a massive refactor to ensure the experience is fluid. Running local LLMs and TTS (Text-to-Speech) engines is resource-intensive, and our recent commits focus on making that process invisible to the user.

Asynchronous Initialization & Splash Screen

One of the most significant changes is how Babilo starts. We’ve moved the LLM and TTS initialization to a dedicated background thread in the Rust backend.

Previously, the UI could hang while the engines loaded. Now, we’ve implemented a two-phase boot sequence:

1. Splash Screen: A lightweight “shimmer” loading bar replaced the old spinner for a more modern feel.
2. Event-Driven Ready State: The backend now emits babilo://core-ready or babilo://core-error events. The frontend waits for these signals before allowing the user to dive into a session.

Centralizing Engine Ownership (The SessionManager)

As the project grows, managing the lifecycle of the AI engines became complex. We’ve refactored the backend to centralize everything within a SessionManager.

• Ownership: AppState no longer holds loose engines; SessionManager is now the single authority, holding engines as internal Arcs.
• Safety: By moving run_turn_streaming() to a synchronous fire-and-forget pattern, we’ve avoided common Rust MutexGuard issues across await points, making the inference loop much more robust.
• Decoupling: Prompt construction is now handled at the session layer (build_turn_prompt), separating the pedagogical logic from the raw inference engine.

Refining the Session UX

We’ve also cleaned up the interface to keep the focus on learning. The biggest change is the Transcript Panel:

• Slide-out Side Panel: Instead of cluttering the main “stage,” the transcript now lives in an animated side panel with a push transition.
• Cleaner Controls: We replaced the old settings button with a dedicated transcript toggle, simplifying the layout to a single-column focus during active practice.

Technical Fixes of Note

• UTF-8 Boundaries: Fixed a bug where the lookahead buffer could flush in the middle of a multi-byte character, ensuring text streaming is always clean.
• Inference Speed: Increased max_output_tokens from 150 to 1000 to allow for longer, more detailed explanations when the mode requires it.
• Race Conditions: We now register stream listeners before calling the backend to ensure no data is lost during the initial milliseconds of an AI response.

Babilo is feeling more like a polished product every day.

See you in the next update!