Gowshik S
All work

Rio

A voice-first desktop agent. Speak once and Rio plans the task, acts on your screen and in Google Workspace, checks the result and reports back by voice.

Role
Solo
Timeline
February to March 2026
Built for
Google Gemini Live Agent Challenge

Every AI assistant today is still a text box with better autocomplete. You still have to run the multi-step task yourself.

One spoken sentence, four stages

  1. Listen

    Live voice in and out, with barge-in so you can interrupt. Push-to-talk, voice activity detection and the wake word “Hey Rio”.

  2. See

    Screen capture and OCR, with Gemini Computer Use for clicks that land on the right pixel.

  3. Act

    Browser control through Playwright. Gmail, Drive, Calendar, Sheets and Docs. Files, shell, windows and clipboard on the local machine.

  4. Check and report

    Rio verifies the result of each step and tells you what happened, by voice.

What is inside

Tools
58
Tools the live voice model calls directly
20
Sub-agents
4: code, creative, research, task executor
Relay rate limit
30 requests a minute, 4 degradation levels
Loop cap
50 iterations
Releases on PyPI
v0.9.0 to v0.9.16

Decisions I would defend in an interview

Built with

Runtime
Python 3.11, FastAPI, Google ADK, google-genai, Vertex AI
Models
Gemini 2.5 Flash Native Audio, Gemini 3 Flash, Gemini 3.1 Pro, Gemini 2.5 Pro
Automation
Playwright, RapidOCR, Gemini Computer Use
Memory
ChromaDB, SQLite
Infrastructure
Docker, Cloud Run, Secret Manager