Finding SaaS Ideas in your Weird Hobbies
A vibe coding story
Coding as a vibe
I’m a lapsed software engineer who became a product manager. I haven’t written a line of code in 20 years. Vibe coding is a new process where non-technical people are creating working software by prompting LLMs like Claude or ChatGPT. My technical background and two decades of building software as a PM made me intrigued by this phenomenon, so I decided to give it a try.
My first few vibe coding projects were just to get a feel for the tools, and understand what their capabilities are. While I am not an amateur when it comes to building software, it’s been over 20 years since I have coded and deployed a piece of software that was used by actual people. The products I have managed have been used by millions of people across several industries, but these were all a team effort involving software and QA engineers, designers, marketers, sales and deployment professionals, technical writer and trainers.
Getting a feel for the tools meant selecting an IDE (Windsurf), picking a Tech Stack (Next.js, TypeScript, Netlify, and Auth0 for auth), selecting my favorite LLM and Agent (Claude 3.7 with thinking). Windsurf is a development environment that connects to your preferred LLM and lets you write code in plain language. The agent interprets your intent, generates code, runs tests, and iterates as needed until you’re satisfied.
The key to keeping the AI from going off the rails is to write detailed PRDs and specifications before hand. Create a simple project plan and to do list, and have the Agent refer back to it before every task. Windsurf has this planning function built in and will take your PRD or prompt, divide it up into steps and then track those steps as it writes and tests the code, which is a huge plus. But, there are many other tools out there which I intend to try on this journey. For example, I am experimenting with Kilo Code, which is an extension that indexes your entire codebase, so that the LLM doesn’t have to spend resources searching the code for specific sections. It also can use the cheaper models, like DeepSeek for planning and architecture and then switch to a more expensive model like Claude to actually generate and test the code.
My first real project
The hamsters running in my head are constantly coming up with ideas, occasionally good ones. But, for my first real project, I wanted to select something that I was uniquely qualified to do.
My weird hobby is that I like to build and race radio controlled cars. Specifically, 1/8 and 1/10th scale off-road cars. I first got into this hobby when I was in middle school, where I saved my allowance and money from doing odd jobs, to purchase my first car.
On the surface, this hobby is completely ridiculous. It is a bunch of grown, middle-aged men refusing to let go of their youth to keep playing with toy cars. However, these toy cars cost several thousand dollars a piece, can reach speeds of 80mph, driving around a tiny racetrack. These races push the human body to the limits of reaction time and hand-eye coordination, while requiring drivers to be extremely meticulous in tuning and setting up their cars. Significant changes to the vehicle setup are measured in 1/2 millimeters and grams of weight make a significant difference in the speed, performance and handling of the cars.
The problem that I am solving is that all of the knowledge for this hobby lives in the brains of 50 year old men. This makes becoming a part of the hobby challenging, as newcomers need to convince a cliquish group of old men to essentially train them on how to get the most out of their cars and go fast.
There are a few books, including the bible but they are, well, books (and insanely expensive). While they may contain all of the necessary information they often won’t be able to answer your specific question. There are also online forums, which are notoriously difficult to search, and you run into the same problem of often not being able to find your exact problem.
This sounds like an obvious use case of an LLM, since it can be trained on obscure data like this and then generate accurate answers to every possible question.
How it works
Each car has something called a set-up sheet, which is a visual representation of all of the possible changes that one could make to the chassis. Just like in real cars, changing geometry, spring rates, shock oils, sway bars, ride height and myriad other settings results in different performance characteristics. In Camberbot, the user selects their car, fills out their setup-sheet online and then can send it to the LLM. The LLM is prompted to review the setup sheet and the user’s question, and suggest recommended changes based on the knowledge contained in its training data.
The LLM wasn’t fine-tuned, but the prompt engineering was deliberate. It was given a set of rules about how to interpret setup sheets, how to think about driver complaints, and it was fed many, many examples to work from. Here is a video demo, or you can try yourself at camberbot.com
I estimate that the total possible user base, globally is less than 1 million people. The hobby is very tiny. I am already in discussions with a few RC retailers to white label this product, or add product recommendations to the result, so a small, but real potential is there. I have a few more things to add like a more robust config screen (to put all of the vehicles on the market) as well as fix a few bugs that I just noticed during the demo, oops. But this is a fully functional MVP that I built, ready to test. It took me approximately 40 hours of work, with about 1/3 the time writing the PRD and Specs and then 2/3 watching the machine write the code, run tests and fix errors.
Stay tuned because my next project will be even bigger. I plan to build a DeFi Liquidity Pool Position Management App with built-in Agentic AI trading. This goal is very ambitious, but I think it can be done. Security and precision will be a much bigger concern, and I will likely need help in this regard.
If you’re into vibe coding, weird side projects, or just want to watch someone build serious apps without writing a single line of code, follow along. I’m going to keep sharing everything I learn. If you want to try Camberbot, send me a message. And if you’re into DeFi or security and want to collaborate, let’s talk.