Ava: Building agentic workflows with Gemini 2.5 Flash and the Live API

Ava is an “AI-powered family operating system” designed to manage the logistics of family life by anticipating needs and automating tasks.

The information parents manage is rarely structured; it arrives via inconsistent school emails, screenshots of flyers, PDF attachments, lengthy WhatsApp threads, and voice notes. Ava must understand the context and seamlessly interact with external services.

To handle the messy, unstructured inputs of the real world, the Ava team implemented a tiered architecture using Gemini 2.5 Flash models for different stages of their agentic pipeline and the Live API to provide a conversational interface.

Incoming requests first encounter a lightweight agent router to make the user experience feel responsive. This router acts as the triage system, classifying the priority of the input, extracting key slots (who, when, where), and deciding which specialized tool or subsequent model is required.

According to Joe Alicata, co-founder and CTO of Ava, “Gemini 2.5 Flash-Lite shines for ultra-lightweight checks,” handling intent detection and short-form summarization while delivering sub-second responses.

Once intent is established, tasks often require deeper reasoning. For example, parsing a school calendar, normalizing inconsistent dates, and proposing the correct event requires nuanced understanding. Gemini 2.5 Flash enables Ava to serve as a capable “household COO” by meeting exacting technical requirements:

• Multimodal understanding: Processing text, images, and audio in a single pass
• Increased accuracy under ambiguity: Correctly interpreting inconsistent school communications
• Reliable function calling: Ensuring that actions, such as calling Gmail and the Calendar API, use structured and trustworthy data

Families can manage their household tasks entirely through voice interactions enabled by the Live API. Alicata noted they had a “hard requirement around native audio” so Ava feels like a natural tool to leverage.

The team used Google AI Studio extensively during development to rapidly iterate on prompts and tool schemas as well as A/B test candidate models, shortening the idea-to-test loop from days to hours.

The results demonstrated the efficacy of their multi-model approach. They observed higher first-pass accuracy on noisy inputs like email threads and photos of flyers. During its alpha sprint, 80% of Ava users were daily active users, and thousands of triaged events were approved and added to calendars.

By using highly-efficient models for fast reads and reserving more resource-intensive models for complex analysis, agentic systems can work at the speed of real life.

To explore how Gemini models and the Live API can streamline agentic workflows, review our API documentation.