February 15, 2024 Recapping the Topcoder Innovation Challenge – IntelliCase Bot (GPT)

Topcoder’s first Innovation Challenge, IntelliCase Bot (GPT), attracted brilliant AI experts from the Topcoder community to push the possibilities of conversational chatbots. Harnessing the capabilities of Large Language Models (LLMs), while wrangling realistically diverse and messy data, participants undertook the exciting task of creating an intelligent assistant capable of navigating and delivering Topcoder data inquires with unprecedented precision.

Topcoder’s Innovation Challenges are a series of AI code competitions designed to ignite innovative solutions at an accelerated pace.

IntelliCase Bot (GPT)

Dec 14th 2023 – Jan 4th 2024
The first challenge invited our experts to develop an Advanced GPT-Powered Conversational Agent that can search and retrieve case studies based on user inquiries (even without perfect keyword match), provide insightful responses with personalized recommendations, all while preserving data location and privacy through an on-premise design.

Even during the holiday season, with an aggressive timeline, our experts provided impressive results that exceeded the ask of this challenge.

Highlight Reel

Intuitive Conversations:

• Chatbots were crafted to understand user queries flawlessly, offering insightful responses beyond mere keyword matching.
• Users enjoyed a seamless experience as the chatbots adeptly handled nuances in their requests.

Privacy-First Approach:

• Embraced open-source LLMs, addressing concerns about data privacy and empowering users with local ownership.
• Process ownership and message control were central, reflecting our commitment to user values.

Mastering Data Chaos:

• Participants tackled the challenge of handling diverse data formats, from PDFs to presentations, with finesse.
• They had to effectively separate signal data from irrelevant noise, ensuring accurate and reliable results.

Real-world Validation:

• Winning solutions underwent rigorous testing on AWS spot instances, validating their viability and response accuracy.
• Final rankings were based on real-world performance, guaranteeing reliability in practical scenarios.

Meet the Champion – Retrieval-Augmented Generation (RAG):

• This cutting-edge framework combined the prowess of LLMs with external knowledge sources.
• Contestants used RAG to enhance LLM responses and provided informed and accurate responses by referencing real-time data.

Tech Choices Simplified:

• Participants explored Mistral and Llama, with Mistral’s Mixture of Experts (MOE) design emerging as the preferred choice.
• The top solution leveraged Mistral 8x7b Instruct, setting a new standard for state-of-the-art LLMs.

User-Friendly Interfaces:

• Chatbot interfaces were designed with simplicity in mind, with options like Gradio, Flask, and Jupyter notebooks catering to diverse user preferences.

Topcoder AI Solutions

Want to integrate GenAI into your product? Need to prove the efficacy of your algorithm? Now’s the time to put your idea in front of the world’s most talented experts at Topcoder.

• Sponsor an Innovation Challenge
• Copy one of ours with your details and data