Rouvia | Devpost

landing page (1)
landing page (2)
Auth0 Sign In
Chat Mode
Sidequest Mode
Profile Details

Inspiration

The inspiration for Rouvia came from a fundamental problem we all face: navigation systems are not designed for how humans actually communicate. This impacts two core areas of travel: safety and experience.

Traditional GPS systems force drivers to manually input destinations while driving, creating dangerous distractions. They require rigid, specific commands instead of understanding natural language creating unsafe and inflexible systems.

But also, we were inspired by the vision of making navigation human again in the way it prioritises the user’s experience. What if you could simply say "I need to get a haircut before my date at 5"? What if planning a trip could be as natural as talking to a friend who knows the area?

Core Features:

Smart Map & Route Planning

Maps your desired experience to a route
Context-aware intent parsing using Google Gemini LLM
Support for complex, multi-destination requests

Unstructured Data Processing

Processed messy voice input and natural language into structured navigation data.
Built a trendiness scoring system using Cohere's AI knowledge to evaluate location popularity.
Created a hybrid pipeline that handles incomplete speech, ambiguous references, and personal context while maintaining data quality through confidence scoring and fallback mechanisms.
Multi-stage processing pipeline: Raw audio → Whisper transcription → Gemini intent parsing → Cohere - RAG personal location matching → trendiness scoring.
Used confidence-based filtering at each stage with structured JSON outputs.
Implemented intelligent caching and distance-based relevance filtering.
Planned web scraping integration for real-time event and activity data from multiple sources

Voice-First Navigation

Natural language processing for conversational requests
Voice-to-text using OpenAI Whisper for accurate transcription
Provides safe hands-free control for drivers

Intelligent Route Optimization

Dynamic constraints (time, distance, preferences)
Real-time route optimization using Google's advanced APIs
Integration with MongoDB to remember saved places and user preferences

Sidequest Mode

Generates spontaneous route suggestions when you have no specific plans
Considers energy level, price, distance, and user interests

Personalized Experience

Learns from user behavior patterns and adapts to past choices
Account registration to save location ie. Home, School, Work etc.

How we built it

Rouvia is built as a modern full-stack application with a sophisticated AI-powered pipeline that transforms voice input into optimized routes.

Technology Stack

Frontend (Next.js 15)

React, TypeScript, and Tailwind CSS
Mapbox GL JS for interactive maps and route visualization
Web Speech API for browser-based voice input

Backend (FastAPI)

Python 3.11 runtime environment
Pydantic for data validation and serialization
Uvicorn ASGI server for high performance

External AI Services

OpenAI Whisper for speech-to-text transcription
Google Gemini for large language model intent parsing
Cohere for trend analysis and content enhancement
Google Places API for location and business information
Google Directions API for route calculation and optimization

Infrastructure

Docker containerization for consistent deployment
Nginx reverse proxy and load balancing
MongoDB for user data and preferences storage
Redis for caching and session management

Multi-Stage Processing Pipelines

Regular Route Planning Pipeline

Raw Audio → Whisper Transcription → Gemini Ambiguous Word Detection → Cohere RAG Personal Location Matching → Complete Order Parsing → Mixed Location Resolution (Personal + Google Places) → Route Optimization → Mapbox Visualization

Sidequest pipeline:

Input parameters → Gemini Intent Parsing → Activity Discovery (Google Places + Eventbrite + Luma + Blog Scraping) → Cohere Trendiness Scoring → Complete Order Parsing → Route Optimization → Mapbox Visualization

Challenges we ran into

Data Quality & Consistency

Challenge: Integrating data from multiple sources (Google Places, Eventbrite, Luma, blogs) with varying quality and formats
Solution: Started smaller, and developed comprehensive data validation, normalization, and enhancement pipelines using Cohere for content analysis

Coordinating multiple AI tools and APIs -Determining the most efficient process for API call sequences

Accomplishments that we're proud of

Turned messy, half-finished voice commands into clear, structured navigation—something that actually feels natural to use.
Made AI personal: the system remembers your own saved places and gets smarter the more you use it.
Designed a trendiness score approach for ever-changing discovery in a diverse world.
We resourcefully pulled together Whisper, Gemini, Cohere, and Mapbox and other APIs into one smooth pipeline, as if it were an agentic workflow, but across tools.
We handled the unpredictable human nuances with confidence scoring, caching, and fallbacks so the system feels reliable, not random.
We set things up for the future: real-time events, social recommendations, and friend-based discovery.

What we learned

How to make the most of different AI models that could communicate a single pipeline
That handling messy, human input is harder than just parsing text, it takes careful confidence checks and fallbacks as there are nearly infinite edge cases.
Personal context transforms navigation from generic to actually useful. In itself it is an endlessly growing dataset, but what matters is being open to understanding various ways to interpret everyday pieces of data.

What's next for Rouvia

Multi-Source Integration: Implementing broader web scraping to Luma, Eventbrite, local Blogs to increase traction of event recommendations.
Mobile App Development: Create native mobile applications for iOS and Android
Social Network RAG: Link knowledge bases across user's friends and social networks, enabling collaborative discovery and friend-based recommendations.