Google Showcases AI-Native App Architecture Paradigm via Agent Platform

Doist built its 'Ramble' feature using Google's Gemini Enterprise Agent Platform (and its predecessor Vertex AI) with Gemini Flash models. The core is Gemini's Live API, which processes raw PCM audio in a single pass for language detection, speech recognition, and semantic understanding, then proactively invokes predefined tools (e.g., addTask).

The architecture employs a layered design with a provider-agnostic streaming layer, a dictation module, the Ramble core module, and a conversation module. This enables rapid rollout of new voice features and flexibility to switch underlying AI providers. For testing, the team combined structural validation with semantic validation using an LLM-as-judge approach, establishing pass-rate thresholds across multilingual scenarios to systematically evaluate model versions.