How I built a multi-agent AI system that automatically finds and validates deals across the web using ensemble modeling, RAG, and agentic workflows.

Click here to watch the demo on YouTube - Speed up the video speed to watch the logging

The Challenge

In today's overwhelming digital marketplace, finding genuine deals feels like searching for a needle in a haystack. Deal aggregation sites are flooded with mediocre offers, and manually tracking prices across multiple sources is time-consuming and inefficient. I wanted to solve this with AI.

The core challenge: Build an autonomous system that can discover deals from RSS feeds, accurately estimate their true market value, and alert users only when genuinely good opportunities arise.

The Solution Architecture

I developed a multi-agent AI system called "The Price is Right" that combines several cutting-edge techniques:

Agent Framework

Scanner Agent: Monitors RSS feeds and identifies promising deals
Specialist Agent: Uses a fine-tuned Qwen model for price estimation
Frontier Agent: Leverages RAG with GPT-4o-mini for contextual pricing
Random Forest Agent: Traditional ML approach with transformer embeddings
Ensemble Agent: Optimally combines all three pricing approaches
Planning Agent: Orchestrates the entire workflow
Messaging Agent: Handles push notifications via Pushover

Agent Architecture

Data Pipeline

RSS Scraping: Continuously monitors deal feeds
Deal Filtering: AI-powered selection of high-quality deals
Price Estimation: Multi-model ensemble for accurate pricing
Opportunity Evaluation: Calculates potential savings
User Notification: Push alerts for significant deals

Building It: A 5-Day Journey

Day 1: Foundation and Model Fine-Tuning

I started with data preparation, curating a dataset of 25,000 home appliance products from Amazon via HuggingFace. The key was creating consistent, high-quality training data:

Data Curation: Standardized product descriptions to 180 tokens to balance information density with training efficiency
Data Cleaning: Removed noise like "Batteries Included," product codes, and other non-valuable text that would waste tokens
Visualization: Analyzed price distributions to understand the dataset's characteristics

For infrastructure, I chose Modal for its serverless AI deployment capabilities. The platform's "code as infrastructure" approach made deploying my fine-tuned model incredibly straightforward.

The fine-tuning results were impressive: using Google Colab's A100 GPU with a batch size of 16, I fine-tuned the Qwen model on all 25,000 datapoints in just 20 minutes. The performance jump from T4 to A100 was remarkable!

Day 2: RAG System and Ensemble Modeling

I built the foundation for contextual price estimation:

Vector Database: Created a Chroma database with SentenceTransformer embeddings for semantic similarity search
RAG Implementation: Built a system using GPT-4o-mini that finds similar products and estimates prices based on comparable items
Ensemble Design: Trained a Random Forest model and created the ensemble logic to combine three approaches: fine-tuned Qwen, RAG-based pricing, and traditional ML

The ensemble approach proved crucial - each method had different strengths that complemented each other beautifully.

Day 3: Automated Deal Discovery

This day focused on creating the autonomous discovery pipeline:

RSS Integration: Connected multiple deal RSS feeds for continuous monitoring
Deal Intelligence: Used GPT-4o-mini with structured outputs to identify the most promising opportunities
Memory Management: Built a system to avoid reprocessing previously analyzed deals

The structured output capability of GPT-4o-mini was perfect for consistently extracting deal information from varied RSS formats.

Day 4: Orchestration and Notifications

I completed the agentic workflow:

Planning Agent: The "brain" that coordinates all other agents and makes decisions about when to alert users
Messaging Agent: Integrated Pushover for real-time push notifications to mobile devices
Workflow Management: Built the logic for seamless coordination between all agents

Day 5: User Interface and Real-Time Monitoring

The final piece was making the system user-friendly:

Gradio Dashboard: Created a web interface showing discovered deals in real-time
Live Logging: Users can watch the agentic workflow in action
Timer-Based Execution: Continuous deal discovery process with user control

Gradio Dashboard

Key Technical Innovations

1. Multi-Model Ensemble Approach

Instead of relying on a single AI model, I combined three distinct approaches:

Fine-tuned LLM: Domain-specific knowledge from 25,000 appliance products
RAG System: Contextual similarity matching for accurate comparisons
Traditional ML: Random Forest with transformer embeddings for baseline accuracy

A pre-trained linear regression model computes the optimal weighted combination for the final price estimate.

2. Agentic Architecture

Each agent has a specific responsibility, making the system:

Maintainable: Easy to update individual components
Extensible: Simple to add new agents or capabilities
Scalable: Can handle thousands of deals across multiple sources

3. Semantic Understanding

Using transformer embeddings enables the system to understand that a "KitchenAid Stand Mixer" and a "Professional Kitchen Mixer" are fundamentally similar products, even with different descriptions.

Infrastructure and Deployment

Modal's serverless platform eliminated the complexity of cloud deployment. The "code as infrastructure" approach meant I could define my entire server configuration in Python.

Chroma Vector Database

Perfect for semantic search and similarity matching, enabling the RAG system to find truly comparable products rather than just keyword matches.

Real-Time Integration

Gradio: Quick web interface development
Pushover: Reliable mobile push notifications
Continuous Processing: Timer-based system for autonomous operation

Results and Impact

The system successfully demonstrates several key capabilities:

Autonomous Operation

Runs continuously without human intervention, monitoring multiple deal sources and making intelligent decisions about what constitutes a genuine opportunity.

Accurate Price Estimation

The multi-model ensemble consistently provides more reliable market value estimates than any individual approach.

Real-Time User Value

Users receive immediate notifications for significant deals, with false positives minimized through intelligent filtering.

Scalable Architecture

The modular design can easily accommodate new product categories, deal sources, or pricing strategies.

Lessons Learned

This project reinforced several important principles:

Ensemble Methods Win: Combining multiple AI approaches consistently outperforms individual models
RAG Adds Context: Similar product matching significantly improves price estimation accuracy
Agent Design Scales: Modular architecture makes complex systems maintainable
Infrastructure Matters: Proper deployment and monitoring are crucial for production AI
User Experience is Everything: Even sophisticated AI needs intuitive interfaces to provide value

Technical Challenges and Solutions

Data Quality

The biggest challenge was ensuring consistent, high-quality training data. Product descriptions varied wildly in length and quality, requiring careful preprocessing and token management.

Model Coordination

Managing multiple AI models and ensuring they work together seamlessly required careful orchestration logic and error handling.

Real-Time Performance

Balancing accuracy with speed meant optimizing model inference times and implementing efficient caching strategies.

Future Enhancements

The modular architecture opens up exciting possibilities:

Multi-Category Support: Expanding beyond appliances to electronics, clothing, etc.
Price Trend Analysis: Historical price tracking for better deal validation
User Personalization: Learning individual preferences and deal thresholds
Social Features: Community deal validation and sharing

Conclusion

"The Price is Right" represents more than just a capstone project—it's a complete AI engineering solution that demonstrates how modern LLM techniques can create genuine real-world value.

The project successfully bridges academic AI research and practical applications, showing how ensemble methods, agentic workflows, and proper infrastructure can create systems that help users save money while requiring minimal manual intervention.

This autonomous deal discovery system validates the power of combining multiple AI techniques: fine-tuned LLMs, RAG, ensemble modeling, and agentic workflows. It proves that with the right architecture, AI can move beyond simple question-answering to create sophisticated systems that actively work on users' behalf.

The 5-day build timeline demonstrates how rapidly powerful AI applications can be developed when leveraging the right tools and techniques. From data curation to deployment, every step reinforced the importance of thoughtful design, proper tooling, and user-focused development.

As AI continues to evolve, projects like this show the immense potential for creating autonomous systems that genuinely improve users' lives—one great deal at a time.

Technical Stack Summary

AI Models: Qwen (fine-tuned), GPT-4o-mini, SentenceTransformers, Random Forest

Infrastructure: Modal, Google Colab A100

Data: Chroma Vector Database, HuggingFace Datasets

Interface: Gradio, Pushover

Languages: Python

Key Libraries: transformers, chromadb, gradio, modal, sentence-transformers

Sources