AI+Science Hackathon

The Data Science Institute and the Schmidt AI in Science Postdoctoral Fellows program are excited to invite you to the upcoming AI in Science Hackathon at the University of Chicago, where machine learning meets scientific discovery! This event is designed for UChicago undergraduates and graduate students eager to tackle real-world scientific challenges using AI. 

The hackathon will take place April 14 – 28 on the University of Chicago campus.

The Kickoff session is on April 14 10:00am – 1:00pm and the final presentations and prize ceremony is on April 28 10:00am – 1:30pm.

Apply soon to secure your spot!  We can only admit a limited number of teams. The deadline is March 13.

Hackathon Highlights

• Two Projects: Competitors will have the opportunity to work on one of two projects. The first project will focus on Leveraging Multi-Agent Frameworks to Extract Clinical Information from Medical Transcripts, and the second project will focus on Predicting Chemical Reaction Structures with Generative AI.
• Team Competition: Each team will have 4 members. Participants can sign up as individuals or as groups of up to 4. Individuals or groups of less than 4 will be matched with others.
• Mentorship: Each team will be assigned a mentor from the Schmidt AI+Science fellows program to provide guidance and support throughout the hackathon.
• Resources: We provide access to NVIDIA A100 GPUs on RCC’s midway cluster to train your models.
• Prize: The winning team of each project will receive a prize.
• Location: Teams will meet in person at the kickoff and final presentation. To get the most out of the Hackathon we highly recommend meeting in person during the Hackathon. We provide rooms on select dates on the UChicago campus to meet and a slack space to coordinate virtually.

First Project Overview

Leveraging Multi-Agent Frameworks to Extract Clinical Information from Medical Transcripts

This project presents an exciting opportunity for participants to use state-of-the-art artificial intelligence methods to extract meaningful information from transcripts of doctor-patient conversations. These transcripts are rich with key pieces of clinical information. However, extracting this information from the transcripts themselves poses a significant challenge. Knowledge graphs (KGs) offer a way to unlock this data: by structuring clinical concepts and their relationships into queryable representations. Recently, healthcare systems have begun to implement Natural Language Processing (NLP) methods to tackle this problem. Typically, this is done using a single-pass query to a powerful large language model (LLM). While this approach shows some success, it is costly and computationally inefficient–making large scale-deployment difficult. Open-source LLMs present a compelling alternative to expensive LLMs. They can be deployed locally (ideal for sensitive patient information) and are cost effective. However, these open-source models face inherent limitations  such as short context windows and insufficient single-pass accuracy for complex tasks. 

For this project, teams will design multi-agent systems to automatically extract structured knowledge graphs from real de-identified doctor-patient transcripts. By decomposing the extraction task across specialized agents with iterative feedback loops, teams can achieve accuracy that rivals commercial models at a fraction of the cost. The challenge is to balance extraction quality with computational efficiency. This means designing agent pipelines that produce accurate, comprehensive knowledge graphs without excessive token consumption or processing time. This challenge offers hands-on experience with LLM orchestration, prompt engineering, knowledge representation, and evaluation-driven system design. All of these skills are crucial for anyone seeking to deploy AI in the future.

Second Project Overview

Predicting Chemical Reaction Structures with Generative AI

This project explores how artificial intelligence can help solve one of the most challenging problems in chemistry: identifying 3D transition state structures that control how chemical reactions proceed. Transition states determine reaction rates and mechanisms, but finding them typically requires expensive quantum calculations and significant expert effort, making large-scale reaction exploration difficult. Advances in generative AI provide new ways to learn chemical transformations directly from data. Instead of relying on traditional trial-and-error simulations, these AI models can be trained to recognize patterns in reactions and propose physically meaningful transition state structures. These ideas open new possibilities for accelerating reaction discovery in catalysis, materials science, and chemical synthesis. 

For this project, participants will work with reaction datasets and molecular structures and develop AI models that learn how molecules transform during reactions. The goal is to explore how modern machine learning can make reaction modeling faster, more scalable, and more accessible.

Eligibility

• Open to all University of Chicago undergraduate and graduate students. If you are a staff member or postdoctoral fellow and are interested in participating, please contact us directly.

How to Apply

• Please fill out the application form. Every team member needs to sign up individually. We have a limited number of spots available, so sign up as soon as possible!

Application Deadline

• March 13