Scenes from the Stanford CodeX Hackathon, April 6th, 2025
Last month, I had the privilege of attending two amazing hackathons here in the San Francisco Bay Area. I was a mentor at the Stanford Law CodeX hackathon that kicked off the Stanford FutureLaw 2025 conference (its 20th anniversary!) and a judge at the PearVC x Anthropic hackathon. Both hackathons were filled with brilliant students, entrepreneurs, and technologists hacking together to conceive, implement, and showcase cutting edge applications built entirely in just 8-10 hours in a single day.
Looking back on these two hackathons, two things stand out the most:
The pace of AI innovation is blistering. Even for those of us that work in it day in and day out, it’s easy to overlook how fast things are moving. We’re still in the early innings and it is increasingly unlikely that there will be any major digital workflows or use-cases that don’t get swept up in the wave.
The human ↔ AI interface matters more than ever. We are still far away from AI agents replacing humans in most non-trivial tasks. They’re still too error prone and fragile. The most impactful AI applications are the ones that that make it effortless and instantaneous for humans to automate the busy work.
In this post, we recap some of the major takeaways and demonstrations from these hackathons and talk about the implications for lawyers and other professional service providers.
Scenes from the PearVC x Anthropic Hackathon, April 26th, 2025. Credit: Katie Li.
Vibe coding is the new hotness
Andrej Karpathy, former director of AI at Tesla and founding member of OpenAI, coined the term “vibe coding” earlier this year to refer to LLM-based coding where users build entire applications entirely through conversing with an LLM and letting the model write, debug, and modify the code with feedback from the user.
If I had to guess, I suspect that at least 90% of the projects in these hackathons were “vibe coded” - because why wouldn’t they be? The speed at which you can get to a fully functional MVP (minimum viable product) is stunning using state-of-the-art tools like Cursor, Claude Code, Lovable, or Bolt. Thanks to the ingenuity and energy of the participants, many of the projects included fully fledged AI agents and even incorporated interfaces like voice or video - incredibly impressive in just a few hours!
Perhaps as a sign that vibe coding doesn’t lead to all sunshine and rainbows, a non-trivial number of projects at the hackathons were dedicated to solving some of the problems created by vibe coding. How do you debug this kind of software? How do you protect it from going off the rails? It’s a microcosm of the AI revolution in a nutshell: there are enormous productivity gains to be had, as long as we can corral all of the new problems that get created along the way.
Production apps dealing with sensitive and regulated data, like in legal, are unlikely to be vibe-coded any time soon, but the line is also going to get blurrier and blurrier. It’s easier than ever to make a flashy UI or an impressive prototype, but it’s important to remember that essential functionality like security, resilience, and scalability are still largely unsolved in the vibe coded world.
One of the PearVC x Anthropic winning projects
My favorite project and what I loved about it
My favorite project from the PearVC x Anthropic hackathon was Ply, by Baladhurgesh Balagurusamy Paramasivan and Barathwaj Anandan. This project won the “Most ready to ship” category at the hackathon.
Ply is simple and elegant: it is a browser extension that autocompletes any form that you are filling inside your browser. It learns about you and remembers that context to fill in the right details later on. It’s an example of how AI is already at a place that can give us superpowers almost anywhere we can imagine. Filling out a job application online? Inputting the same details over and over into sign up forms? Trying to figure out the exact right Google search keywords? AI autocomplete makes the right context available to you at exactly the right time.
This is also an example of how integrating AI deeply into workflows with simple and intuitive interfaces creates fundamentally better experiences than chatbots in many cases. This is an area that I expect we’ll continue to see a lot of innovation in.
What does this have to do with automated timekeeping?
When we started Billables, we wanted to bring AI to bear to automate the most mundane, painful, and repetitive workflows for lawyers and other professional service providers, starting with the most dreaded, “bane of my existence” one of all: timekeeping. And we wanted to do so in a way that felt effortless and intuitive, an obvious extension to the way people work already.
Observing the creativity and breadth of the prototypes that these brilliant teams were able to put together in just a few hours reinforces a few of our core assumptions. In particular, we believe that for AI applications to add unique and differentiated value:
You need to meet users where they are. There are too many tools to learn, and not enough time to try them all. No one wants to spend time becoming a prompt engineer in 2025, and in a crowded market, UI/UX stands out more than ever before.
Quality matters, and quality depends on data. It is very hard nowadays to have a “smarter” model than anyone else, in any particular thing. General purpose models are still improving in leaps and bounds. What matters is the system you put around it: what data it has access to, how you protect that data, and how you combine the right tools at the right times.
The best AI use-cases are the ones that automate the things humans don’t want to do. We’re still early in the AI curve and while AI agents have a lot of potential, there are a lot of things they don’t do well. Today, there is enormous opportunity to automate low risk, high value workflows like timekeeping that the majority of humans want nothing to do with.
Final thoughts
AI tools, and especially the tools we use to build AI applications, are getting better daily. It’s a mistake to underestimate what these applications will be capable of in the future, and rather than fear it, we should all be thinking about how to harness it to free up our valuable time and resources for the most creative, stimulating, and valuable work. Cutting edge AI applications also introduce a host of new problems that still have to be solved, so in my humble opinion, it’s going to be a while before AI agents replace most of us. In the meantime, the key is to build applications (like Billables!) where AI feels like a natural extension of a user, helping them offload all of the painful and repetitive stuff.
Acknowledgements
I would like to extend my sincere gratitude to our advisor, Dr. Megan Ma, for inviting me to be a mentor for the Stanford Law CodeX hackathon, and to Annie Ta and the Pear VC team for having me as a judge for the PearVC x Anthropic hackathon. It was a pleasure to meet so many smart, ambitious and capable participants and to support these two fantastic events.
