What is the problem being addressed?
Our senior design project focuses on improving the pricing intelligence behind Throwly’s AI-powered marketplace. For the platform to be useful and trustworthy, it needs to generate price estimates that are both accurate and fast across a wide variety of items and services. The challenge is that marketplace listings can vary greatly in category, quality, and available context, which makes consistent pricing difficult.
The existing system worked for a limited set of categories, but it needed stronger accuracy, broader coverage, better validation, and a more efficient pipeline. Our team addressed this problem by focusing on how Throwly’s AI estimates value, how that performance can be measured quantitatively, and how the overall system can be improved to better support real-world marketplace use.
What is our project idea and how it works
Our project improves Throwly’s AI pricing pipeline and provides a foundation for future development through research, testing, and system recommendations. The main components of our work include:
- Improving the accuracy and speed of AI-generated price estimates for common marketplace and household items
- Expanding support to more categories, with emphasis on frequently listed items such as fashion and furniture
- Building a diverse validation set and a reusable validation script so future data scientists can measure model performance consistently
- Testing different data science and machine learning approaches, including prompt engineering, model choices, and context strategies
- Adding data-cleaning steps to make the system more robust against low-quality inputs and unsafe or misleading context
- Integrating stronger external context sources, such as marketplace data and internet research, to improve pricing quality
- Supporting service-based pricing so the AI can estimate the value of local services through deeper research
- Reducing unnecessary pipeline inefficiencies to improve response time and overall reliability
- Recommending improvements to the current API pipeline, website, and MCP server architecture
- Summarizing our findings in a research-style report with data, visualizations, methods, and recommendations
The result of this project is not only a set of technical improvements, but also a clearer, more quantitative framework for evaluating and improving Throwly’s AI systems over time.