Collaborative Output-time Neural Networks for Enhanced Reasoning (CONNER)
Inspiration
Our framework draws inspiration from cutting-edge AI reasoning models including Deepseek Prover, Tree-of-Thought methodologies, and Design-by-Contract (DBC) principles.
What it does
CONNER provides transparent, controllable AI reasoning for complex administrative tasks, allowing users to monitor each step of the AI's decision-making process and intervene when necessary.
How we built it
We developed CONNER using a Streamlit front-end for the user interface, Gemini Flash to power MCP tool integration, and OpenAI GPT-4 to handle the main agent reasoning flow.
Challenges we ran into
We couldn't find a Python Agents SDK that supported conversations as input rather than just text, forcing us to implement MCP agents from scratch. Additionally, we encountered numerous Streamlit display bugs and struggled with linking the agent seamlessly to the Streamlit interface.
Accomplishments that we're proud of
We successfully built MCP agents from the ground up and created an innovative node-flow system that allows users to view and modify agent reasoning in real-time.
What we learned
Through this project, we gained deep insights into creating effective AI agents and identified key barriers preventing agentic AI adoption in workplace settings, along with strategies to overcome them.
What's next for C.O.N.N.E.R.
We plan to enhance the node display system by implementing react-flow instead of basic text boxes and will add conversation history functionality to reference previous interactions. We're also working on maintaining version control and allowing users to view complete reasoning trees from past sessions.
Log in or sign up for Devpost to join the conversation.