Hack Your Weekend

There’s a big difference between spinning up an MVP in Lovable with a single prompt and spending an entire weekend working hands-on with Claude Code Max - versioning, MCPs, documentation, staging, app security, and everything in between. In our team, I took care of the backend, deployment, and app tracking (our tech stack was React + Next.js + Supabase Postgres DB + Vercel hosting + Git, plus tracking via a Cloud Run endpoint writing to BigQuery, along with a dashboard and data interpretation through a BigQuery MCP).

The event was organized by Jindrich Faborsky (probably Jindra’s smallest event ever :). Jindrich is really pushing hard into AI-assisted development - he genuinely lives it - and it’s great that he openly shares his journey while onboarding both the community and people from the outside. If you’re looking for an intro to the topic, I can highly recommend Jindrich’s AI First Course - an authentic course about the core principles, the tooling, and how to actually work this way. It was also great that the event brought in mentors from different areas of expertise who supported the teams - people like Petr Bureš, Lukas Mehnert, Dima Melnik, or Petr Simecek, who helped us with Git. Jindra himself actively supported the teams too - helping some with development or strategy and nudging others (like us) toward even bolder features. 🙂

At #HackYourWeekend, our Keshu team (Klára Fottová, Jiří Pavlas, Petra Schönfeldová, Michal Pecánek, Vito Nikolič + me) built an application for listing and searching dog-friendly places - kind of like a European Coffee Trip, but for dogs. (Keshu was the dog/mascot.) My role was backend development and deployment (a dream come true for me - finally getting to try the role of a backend engineer :). The other roles covered design, development, business, and marketing. We designed a nice data model (pro tip: the schema was written in Mermaid, which Git can automatically render from a .md file), and before the rest of the team jumped into the app, I pushed the first 30 commits.

Our tech stack was React + Next.js + Postgres DB (Supabase) + Vercel hosting + Git. On Sunday, once the app was deployed with the basic functionality and admin panel in place - and my teammates were working on new features and design (respect to Michal Pecánek for your fully working geolocation 👌) and I built a custom tracking setup using Cloud Run and BigQuery.

Other teams built, for example, a weather-check app for hobby pilots to use during pre-flight preparation (Briefly.ai) or an app for sharing routes and rides within the cycling community (Bike Meet). The project that caught my attention the most was an app with the working title “Střídavka” (“Shared Custody”), designed to help separated parents organize shared custody - managing handovers, schedule changes, and day-to-day coordination. It’s a powerful topic and an app that truly makes sense and brings real value, and I genuinely hope its development continues.

‍

Interesting Bits from Development & Lessons Learned

We eventually had to switch to Claude Code Max, because the standard PRO version disconnects you after a few hours - not ideal for a hackathon - and it burns through tokens fast. I honestly can’t imagine building a fully robust solution this way yet, at least not without splitting it into well-isolated components so the model can keep the context under control. One of the highlights for me was having the time to really try out a bunch of MCPs. Supabase and Vercel both have their own MCPs, which makes development so much easier, and we also plugged in others (Context7, Sentry, etc.). Honestly, I wouldn’t bet too much on the future of services that don’t have an MCP - and don’t plan to build one.

Version control turned out to be one of our biggest challenges. We didn’t really have a good branch strategy, and we lacked the experience to merge everything properly across a six-person team. What helped me a lot was having Claude connected directly to Git. I committed straight from Claude and had it check in advance whether my commit would break the build. It wasn’t exactly professional - and proper tests should be done differently - but we didn’t have much time to fine-tune things.

One thing that didn’t work very well for me was the project instructions in .claude.md. I wanted Claude to automatically update the prod.md file after every functionality update, but I couldn’t get it to work. It only worked when using an agent - and even then, I had to trigger it manually.

‍

App Tracking

We built our own tracking (GA4 would’ve been way too easy :). We sent events through a secure cloud endpoint on Cloud Run directly into BigQuery. In a real-world scenario, I got to try building a Claude Code endpoint that processes incoming events and writes them to BigQuery - plus hooking up and testing the BigQuery MCP. GCP has something called the MCP Toolbox for Databases, which includes several MCPs for working with data, including the BigQuery MCP. Thanks to that, we were able to prompt our way to full dashboards, and we were even ready for on-the-fly data queries (for now just in the console, but the next step would be wrapping that in a UI and deploying it into the app’s admin panel).

There were folks from Keboola at the event (huge thanks for their support and sponsorship 🙏), and apparently the Keboola MCP is already capable of handling most tasks - including full end-to-end workflows. I didn’t know that, and I’m definitely planning to try it out.

‍

Claude Code + Gemini on the same project

I experimented with combining two models on the same app. I built the tracking layer separately using Gemini CLI and the Google Cloud SDK. I then used Claude Code for deployment, because I didn’t want to deploy from two different CLIs or spin up the Vercel MCP inside Gemini. I was a bit worried whether having two models working on the same codebase would cause issues.

But if the second model is working on a contained component, you avoid major structural changes, and you document everything properly so the other model can always reload the context, I think it’s totally fine.

I documented all changes in tracking/tracking.md and always let Claude review what Gemini did and confirm that everything looked good from its perspective.

‍

What I learned - and what I’m taking away from the hackathon for my work and personal growth

I left the event tired, but happy - and grateful for the chance to take part. I had already tried building with AI before the hackathon (thanks, FOMO), but what I took away from Brno most was the experience of building something as a team. And also getting more comfortable controlling things in Google Cloud Platform directly through Gemini or preparing Cloud Shell commands via the model.

Honestly, I’m not sure I’d feel confident building an app intended to serve a large number of users without some human code review. But there were people at the hackathon who do have ambitions to push their apps further - and thanks to the hackathon, they can now build an MVP themselves and really take off. For me, this approach feels more useful for various smaller data tasks, system integrations, test agents, utilities, pipelines, and similar. Basically for things that aren’t mission-critical but save us a lot of time at MeasureDesign.

It massively democratizes access to technology and removes the barriers created by not knowing a technology in depth - or not having time to build something complex in it. What matters is knowing what the tools can do, what you want from them, and what you need to keep an eye on. What worked well for us was keeping the AI within boundaries, avoiding too much “creativity,” and breaking development into detailed steps for the model to execute. With that approach, the results were genuinely good.

Fun fact to wrap it up: I ended up sleeping in my car 🚘 on the street in front of Clubco both nights, because at 9 p.m. on Friday I realized that my accommodation wasn’t booked from September 19th, but from October 19th 🤦. I even tried sleeping in a meeting room - you know, since it was a hackathon (like my teammate in the photo) - but I couldn’t fall asleep there, so the car won. In the end, it didn’t matter at all - actually the opposite. There wasn’t much time to sleep anyway, let alone head somewhere to check into a hotel. :)