Inspiration
I - Akash - rotated through a couple of different labs in particle physics during my undergrad, and the issue I always ran into was bespoke software & languages that took me time to learn, leaving me little time to do the physics that I really wanted. We wanted to create something that abstracts away this process and allow students & researchers to focus on the physics. Longer term we want to make this into a tool that can be plugged into any frontier model via a shell command, allowing them greater mastery over physics.
What it does
It allows students & researchers to query natural language physics questions to an AI, and instead of guessing or building hypotheticals, the AI is now able to design, execute, and analyze a real experiment built on industry standard tooling. This is like giving the AI access to a lab to perform experiments in.
How we built it
We built two solutions in parallel: one native macOS app and another a extension/plugin for CLI coding agents. Both the solutions took inspiration from each other. But while building the native app we realized that putting a smart agent in a harness actually limited it's capabilities, whereas putting it inside a coding agent allowed it to use the tool we built in ways we didn't even anticipate. We used Swift for native, Python for CLI, and V0 typescript for the frontend.
Challenges we ran into
UI complexity with swift apps, API limits on gemini models, different requirements for different coding models, traceability with agentic actions, tradeoff between allowing external libraries and safety of running generated C++ code raw on a machine, designing sandboxing techniques for said raw C++ code.
Accomplishments that we're proud of
The fact that we were able to get to two working solutions. Watching the native app handle larger simulations with ease, and then watching the CLI version get to a result using the agentic capabilities in ways we didn't anticipate. It was all very exciting.
What we learned
We should focus on one solution going forward. The ideas started with similar sets of tools and functionality but the drift became apparent by day two. Some functionality is native only, some are CLI only. And I think if we focused on one platform - we could have had a stronger or more scoped out solution.
What's next for Vidura Labs Alpha
Reaching out to my old physics professors for a demo run, and potential pilot with his students..
Log in or sign up for Devpost to join the conversation.