Skip to content

Quadra

We built a friendly, intent-aware maps web app using Google AI Studio for rapid iteration and code generation.

Code Demo Cerebral Valley

Core trick

Gemini grounded in Google Maps and Search re-ranks nearby places against a free-text intent and surfaces one top match.

Tags

productivity

Description

We built a friendly, intent-aware maps web app using Google AI Studio for rapid iteration and code generation. Gemini models, grounded in Google Maps and Search, analyze geospatial data to provide tailored, context-aware suggestions. Today’s maps show one-size-fits-all lists, forcing manual filtering and slow decisions. Instead, we capture intent and use context to re-rank results so each person gets the right place for their immediate need. Enter Star Guide. It understands complex, shifting goals like “Wiped out” (close by), “Good Coffee” (decent quality), and “Meltdown” (quiet, changing station). Using Gemini, we skip generic lists and instantly re-rank local options based on how perfect they are for your situation right now, pulling signals from everything. We surface a single best match so you stop sorting through junk and make the right decision instantly.
SonnyDev/https-github.com-SonnyDev-gemini-guide★ 00 forksTypeScriptCreated Nov 15, 2025 · last push Nov 15, 2025

README excerpt

# Gemini City Guide 🇬🇧

An interactive, multimodal tour guide application powered by **Google Gemini 2.5** and **Google AI Studio**.

This application demonstrates the capabilities of the Gemini Live API to create a real-time, voice-interactive agent that not only converses naturally but also actively controls the user interface (Google Maps) to provide a rich, guided experience.

## 🤖 AI Implementation

We leverage the `@google/genai` SDK to integrate Gemini's advanced multimodal features:

### 1. Real-time Voice Interaction (Gemini Live)
The core experience is powered by the `gemini-2.5-flash-native-audio-preview` model via the Live API.
- **Bi-directional Streaming**: We establish a persistent WebSocket connection using `ai.live.connect`.
- **Audio Processing**: The app streams raw PCM audio input from the microphone and decodes the model's audio output using the Web Audio API for low-latency playback.
- **Voice Personalities**: We use `speechConfig` to assign distinct voices (e.g., 'Fenrir', 'Kore', 'Zephyr') to different guide personas, complete with custom system instructions for tone and style.

### 2. Visual Tool Calling (Agentic UI)
The model acts as an agent that can manipulate the UI. We define tools like `update_map` and `get_directions` in the session configuration.
- **Function Execution**: When the model decides to show a location (e.g., "I'll show you where Big Ben is"), it sends a `toolCall` instead of just text.
- **Visual Cursor Animation**: The app interce

Team