Building on your earlier note about being 'custom-built,' could you elaborate on customization? For example, is it a fine-tune of an existing model, and if so, which one, to optimize for low-latency relay interactions?