Data Science Festival kicked off its 10th anniversary year with a hands-on Sandbox Session featuring Jason and myself. The energy was great, the chat was active, and many people built something useful in about an hour.

The Problem To Solve For - EV Charging Stations

Where should a city place EV charging stations to best serve demand? I framed it as an optimization challenge where you model a city as zones with locations and EV charging demand, then figure out where to put a limited number of charging stations so high-demand zones get served efficiently and travel distance to the nearest station stays minimal.

I walked through it step-by-step, breaking the problem into discrete prompts:

• Generate synthetic city data with zones and demand weights

• Compute a distance matrix between all zones

• Define an objective function for facility placement

• Run K-medoids optimization to find the best station locations

• Add a maximum service radius constraint

• Visualize the results

Jason jumped in full send with a single comprehensive prompt and just let the agent run. His version ended up doing way more than the original scope, including temporal demand patterns (when stations would be busiest), risk analysis for what happens if a station goes offline, and comparisons across three different optimization methods.

Both approaches used almost identical credits (23-24) and got to similar core results. My method gave me more control and comprehension of what was happening. Jason's produced richer exploratory analysis but was harder to follow in real-time.

"Honestly, that is so cool. I feel like I've seen the future. Just my head's exploding with possibilities." - David Loughlan, Founder, Data Science Festival

One surprise from Jason's run: the agent discovered you could actually meet all the coverage constraints with just three stations instead of the expected ten or fifteen. The greedy set approach found that configuration automatically.

Top Highlights from the Session

My optimization hit a wall when the agent couldn't find a feasible solution with the initial service radius. The agent recognized the issue, increased the radius from 45 to 47 units based on the output showing the unconstrained solution had a max distance of 46.87, and reran the optimization successfully.

Zerve’s collaborative features got a shoutout too. When I invited Jason into my project mid-session, all the variables were already loaded. Zerve serializes everything after each block runs, so collaborators can jump in and pick up where things left off without re-running any code.

"I love the UI. I love what you've built in terms of the layout - it was super intuitive. As a  first-time user. I'm like, ‘oh, that makes sense!’. Very quickly, I worked my way around, which was nice." - David Loughlan, Founder, Data Science Festival

Several participants created their own variations and shared results in the chat.

Greg & Jason’s Top Zerve Prompting Tips

I always try the simplest possible prompt first, then build from there. He mentioned seeing people spend an hour writing a prompt when they could just try something easy and see what the agent does. 

Jason takes a similar approach but thinks in terms of milestones. He'll tell the agent to get to a certain point, look at what comes back, and then decide whether to push forward or change direction.

Get More Credits

If you want to try this yourself, you can extend your runway by signing up for Pro before the end of January 2026. You will get your first month of Pro free, which includes 250 credits. (This promo ends January 31, 2026)

The session wrapped with plans for a fantasy football hackathon. Given how quickly Jason spun up a Premier League prediction model during the demo, it should be a good one.

Watch the recording here: