Snapr: A Multi-Agent Collaborative Real-Time Deal Discovery System

Snapr is an open-source multi-agent AI system that integrates multi-agent architecture, LLM, traditional machine learning models, and RAG technology to achieve intelligent discovery, filtering, and price prediction of real-time online deals. Its core design philosophy is division of labor and collaboration, where a central coordinator manages various specialized agents to complete complex tasks, featuring modularity and strong scalability.

In the e-commerce era, consumers face the paradox of information overload and difficulty finding high-quality deals, making manual tracking impractical. Traditional price comparison websites can only crawl public prices and lack intelligent evaluation and price prediction capabilities. Deal discovery involves multiple tasks such as real-time data crawling, deduplication filtering, and quality assessment. A single model is difficult to handle all these tasks, and a simple pipeline lacks flexibility—this provides an application scenario for multi-agent systems.

Snapr adopts a central coordinator architecture, encapsulating AI capabilities into specialized agents: the data collection agent is responsible for real-time crawling of e-commerce data; the filtering agent uses embedding vectors and RAG to deduplicate low-quality deals; the prediction agent combines historical data and ML models to predict price trends; the evaluation agent uses LLM to analyze deal value (brand, reviews, historical prices, etc.). This architecture is modular, facilitating component upgrades and expansion.

Technology stack: Python3.11, uv package manager, Modal serverless platform, Docker containerization, GitHub Actions CI/CD, FastAPI service. Innovative applications of RAG and embedding technology: encode historical deals into vector embeddings for storage, find comparable records through similarity search; use semantic similarity to identify duplicate product links and avoid redundant information.

LLM is responsible for natural language understanding and reasoning tasks (such as deal description parsing, review summarization, value evaluation); traditional ML handles structured data prediction (such as price trend, sales volume prediction). The outputs of both are fused at the central coordinator—for example, if ML predicts a price drop and LLM finds an increase in quality complaints, the coordinator makes a comprehensive decision to delay the recommendation, improving decision accuracy.

Snapr follows excellent engineering practices: clear code organization, separation of configuration and code, management of sensitive information using environment variables. The CI/CD process includes pre-commit checks, automated testing, container building, and deployment. Deployment uses the Modal serverless platform (supporting GPU acceleration and automatic scaling), single-container deployment simplifies operation and maintenance, FastAPI serves both Gradio interface and MkDocs documentation, and Docker ensures cross-environment consistency.

Snapr's core capabilities can be adapted to various scenarios: stock price alert, flight fare tracking, rental screening, academic paper recommendation, etc. (other components can be reused by replacing the data collection agent). The project provides the my_ai_portfolio series of experiments (including Jupyter Notebooks) and detailed architecture documentation, offering learners a path from theory to practice.

Snapr represents the trend of AI applications shifting from single models to multi-agent collaboration. Solving complex real-world problems requires a combination of multiple professional capabilities. Its value lies in demonstrating the engineering construction of multi-agent systems (architecture, technology selection, development process, deployment), making it an open-source case worth studying. In the future, as LLM and agent frameworks mature, more intelligent systems that automatically complete complex tasks will emerge.