SolMate

Inspiration

As we witnessed the remarkable progress of agentic AI across various industries, we recognized a significant gap in the Web3 space: the complexity of managing digital assets was becoming a barrier to wider adoption. This insight led to the creation of SolMate, an AI-powered crypto assistant that bridges the gap between traditional user experiences and the crypto ecosystem.

What it does

SolMate serves as your AI-powered crypto guru, seamlessly bridging the gap between blockchain operations and natural human interaction. Here's how it empowers users:

Execute swaps and transfers with conversational commands like "swap all of my Solana into USDC"
Get real-time price checks and market updates through simple questions
Track your entire crypto portfolio through natural conversation
Set up price alerts using natural language (e.g., "Alert me if ETH drops below $2000")
Model has access to real-time market analysis and trends
Get price predictions and market sentiment analysis
Save wallet addresses as aliases for easy reference when interacting with the model

How we built it

Core Intelligence

Leveraged Llama 3.3 Instruct for advanced tool calling and instruction parsing
Integrated Eleven Labs for voice interaction capabilities
Built RAG system using bge-large-en-v1.5 embeddings and MongoDB vector search

Real-Time Market Intelligence

Developed a custom market data pipeline using yfinance for real-time price tracking and market updates
Created a RAG system using bge-large-en-v1.5 for embedding market-related content
Web scraped to gather and process latest crypto news and market analyses
Utilized MongoDB with vector search capabilities for efficient retrieval of market insights

Blockchain Integration

Integrated Phantom wallet SDK for secure transaction handling
Built custom middleware to handle wallet connections and transaction signing

Alert System & Notifications

Developed an SMTP-based alert system for price notifications
Created a persistent monitoring service for tracking user-defined price targets

Challenges we ran into

Working with instruction-tuned models proved to be our biggest challenge, as Llama 3.3 Instruct required heavy prompt engineering to achieve consistent function calling behavior. The model would sometimes ignore available functions, making it paramount to carefully craft our prompts.

Accomplishments that we're proud of

We're particularly proud of piecing together multiple complex systems into a seamless user experience. Despite the challenges of working with so many moving parts, we successfully created a natural language interface that reliably handles crypto operations, combining real-time market data, blockchain interactions, and voice capabilities.

What we learned

Instruct models are extremely pedantic.

What's next for SolMate

The future for SolMate includes expanding on its trading capabilities by integrating it with major DEXs and CEXs to provide a wider range of trade options.