K.I.J.I.J.I. S.P.A.M.M.E.R.

💡 Inspiration

Millions of users access the second-hand market Kijiji daily, but buyers often struggle with finding relevant products and securing the best deals. K.I.J.I.J.I. S.P.A.M.M.E.R. is designed to revolutionize this process by utilizing LLM agents. It aims to automate the tedious tasks of scrolling through listings and haggling with sellers, making the entire experience more efficient.

🫣 What it does

Automates the end-to-end process of finding and negotiating purchases on Kijiji.
Employs comprehensive web scraping techniques to search and filter listings based on the buyer's criteria, ensuring relevant results.
Utilizes LLM-powered smart agents that initiate conversations with sellers, adopting various buyer personalities (casual, urgent, lowball, etc.) each with their unique negotiation style.
These agents work together, strategically pushing towards the buyer's desired price range. Once a target price is reached, the buyer is notified to accept the deal.

👷 How we built it

We combined advanced web scraping technologies with the power of LLMs to gather and structure relevant listing data, including product details, pricing, and seller information.
The core of K.I.J.I.J.I. S.P.A.M.M.E.R. is its smart agents, engineered with distinct personalities and bargaining strategies. These agents analyze conversation history and coordinate their approaches to optimize negotiations.
Through prompt engineering, we created dynamic personalities that can adapt their tactics based on seller responses, ensuring effective negotiations across multiple conversations.

😬 Challenges we ran into

Identifying scope; building unnecessary tooling outside of the final scope of project.
Adapting the Cohere API to facilitate multi-turn conversations and maintain contextual awareness of previous chat histories
Engineering the LLM and its prompts to teach the agent to adopt multiple personas. We had to carefully tailor the prompt for the agent to coordinate actions between its different personas to execute its bargaining strategy.
Optimizing our system to minimize latency and mimic "real-time" messaging as closely as possible.

✨ Accomplishments that we're proud of

Building modular, well-structured code
Integrating automated browsing technologies

🤓 What we learned

Cohere API
Webscraping
Tkinter

🌬️ What's next for K.I.J.I.J.I. S.P.A.M.M.E.R.

Automating Sell-side: Expanding beyond the buy-side, K.I.J.I.J.I. S.P.A.M.M.E.R. will develop tools to automate the selling process as well. Sellers will be able to list their items, and our smart agents will handle inquiries, negotiate with potential buyers, and ensure sellers achieve the best possible price without the hassle of manual responses.
Expanding Marketplaces: We aim to extend K.I.J.I.J.I. S.P.A.M.M.E.R. to other platforms such as Craigslist, Facebook Marketplace, and eBay, making it a versatile tool for any second-hand shopper.
Enhanced Negotiation Algorithms: Continuously refining the personas’ strategies, making them more adaptive and capable of leveraging historical data to improve bargaining success.
User-Controlled Personas: Allow users to customize their agents’ negotiation styles, giving them more control over how aggressive or patient their bargaining process is.
Advanced Price Analytics: Implement machine learning to analyze market trends, giving users insights on when and where to get the best deals based on historical pricing data.