I went to an AI Hackathon, here’s what I learned
From ideation to working demo and pitch in 8 hours, what AI coding can do for you. By JUNC Founder Noah Allen.
CS majors might be cooked.
How did I come to this conclusion? Well, I recently participated in a hackathon/pitch competition with zero backend, frontend, react, java, or flutter experience. I went because the event advertised how to use AI tools and develop projects from scratch. It was important to me to find out exactly what these tools might be capable of and how they could be used. We were encouraged to use an integrated development environment with an agentic AI attached, known as Google AI Studio. Together, this platform enables users to launch demos in under ten minutes. I learned a lot of technical terms and tools (react, java, etc.), but what happened next convinced me I won’t ever need to learn them deeply.
Our task was to utilize open data, preferably from the state of New York, to solve civic problems involving public transportation and childcare. We were given pizza, soda, and five hours to work on our problem. By the end of the day, we needed a working demo and a five-minute pitch deck to compete for cash prizes. This wasn’t my first time pitching, and definitely not my first time pulling a presentation together the same day, but it was the first time I’ve ever coded an entire application.
By 30 minutes in, I was seriously blown away. The Google AI studio allowed for rapid prototyping of a tech stack capable of scraping website data, storing it in a database, and building a live dashboard in about 10 minutes. The real advantage was in the simplicity and speed of the workflow. I could go straight from a research question to visualizing an answer, bypassing the potential struggles of coding. Operationally, this looks like: 1. Find an interesting data set. 2. Develop my research question. 3. Build a prompt including the information from step 1 and 2, then request plots and statistical tests.
This system was working very well, giving great insights, until I realized the AI agent was fabricating all of the data…. File this under reasons AI native users are cooked.
My bioinformatics brain knows when something looks too good to be true, too smooth, or just works too easily. Apparently, the AI agent was biased toward running lightweight dashboard demos, meaning it would look at the structure of the data then guess what the average values would be to make plots that satisfied our hypothesis. After some struggling, and very direct prompt engineering, we coaxed the platform to use REAL data in our plots. I still wonder how many of the other teams were misled by their data being unknowingly fabricated. Regardless, by the end of the five-hour session, 15 out of 15 teams had working prototypes, including maps with predictions of bus station accessibility compliance (my team), home inspection tools, and a bird-watching application.
In the end, this hack-a-thon opened the door for me to learn so much more about new implementations of AI. Up to this point, I’d only ever worked with ChatGPT in a back-and-forth fashion to modify my R codes. After running afoul of the mock data being used in Google AI Studio, I switched to their Antigravity studio system. Antigravity is a fork of VS code built to assist coding with AI agents. This platform has allowed for far more oversight, local execution of scripts, and control over input data. In the week following the hackathon, I wrote four more applications:
1. A dashboard that pulls meta data from the monthly uploads to the NCBI Gene Expression Omnibus (GEO) Repository, tracks trends, identifies the top data sets based on the impact of their associated citation, then writes a report on all the findings.
2. A dashboard that pulls economic indicators of scientific impact from monthly government reports (e.g., STEM-related patents, FDA drug filings, jobs, renewable energy).
3. An app that takes my order request forms, extracts the metadata, writes an email, and asks for my review before sending it for purchasing approval.
4. An iOS/Android app to log board game statistics, analyze trends, and highlight game play strategies.
The first app wrote itself on a Monday while I was doing lab work. I was two floors below my office listening to a podcast and changing cell culture media. There was no back and forth with ChatGPT, searching StackOverflow, or watching YouTube tutorials. I prompted my way to working tools in under an hour. Weirdly, the only trouble now is something like writer’s block, staring at a blank prompt screen wondering what I should create. The key was, and still is, I had a question to pursue. That remains the core of what I do as a scientist, regardless of the tools available to me. I come up with a question, then I use the necessary tools, resources, and relationships to solve it. This process generates something new and hopefully insightful.
I will continue this series over the next several months, documenting my experiences building with new tools and using the Google agentic integrated development environment Antigravity. In the next installment, we will see how data from the NIH Data Book can be used in the search for NIH grant funding. This series, titled “Public Builds” will be accompanied by Github repositories to track and share development along the way.