Table of Contents

• Introduction
• Features
• Technical Stack
  • Challenges Faced
• Future Enhancements

Introduction

AI is pretty much everywhere. There's hardly a week without some kind of AI news. New models with innovative architectures and cheaper inference costs. Pledges for massive datacenter and infrastructure investments. Some new domain that we slapped AI on (sure we've seen AI in healthcare, but have you seen AI in a smartwatch for healthcare?), or simply significant improvements in existing domains like image generation. With these advancements come a slew of "slop" products. For the most part, these are just complicated wrappers over existing AI models, making them even easier and more accessible to use. For example, while there isn't a practical difference between just using ChatGPT directly and using an AI IDE like Cursor, the built-in integration and tooling dramatically alters user experience and expectations. Especially in software engineering, there have been entire new cultures and definitions that have developed specifically as a response to this AI (over) saturation - vibe coding and a new breed of founder, to name a few. Unfortunately, a large number of models (and AI apps) are closed-source, cost a lot of clams, or are simply infeasible to run anywhere except a cloud hosting provider due to their massive size. Not to mention the fairly self evident dangers of AI built and run by a select few billion dollar tech giant monopolies. However, thanks to the work of Ollama and research teams that release their work, it is easy to find decent models that are not only open source but small enough to run on consumer hardware. These emerging technologies and applications have been a massive inspiration and motivation for this project. How can I contribute to this ecosystem of AI apps, using the tools that Ollama provides? How can we improve the experience of developers locally, using their machine? Ultimately, I decided to recreate v0 by Vercel. I thought the design was stunning, and it would be a great way to test my new libraries that I'll describe in further detail later.

Features

Vercel's v0 dashboard

The real v0 has a lot. A bunch of random tabs (community, projects, feedback), ways to manage your account, templates, the actual chat and editor interface, and more. However, my project is merely a clone that's designed for local use. I don't need to implement a community tab or accounts when everything should be offline. Above all, the most important feature is generating a user interface with an LLM, and providing some mechanism to iterate over and copy the resulting code. Being in school has really limited my free programming time, so I focused on the essentials:

• A similarly styled input field
• A way to view what Ollama spits back out at us

v0 (clone) input field with past projects

v0 (clone) generating a UI

Technical Stack

I used the libraries and frameworks I was most familiar with to help speed up the development process. I used Remix, React, TypeScript, and of course Ollama to provide the necessary models and API service. However, v0 was so fun (and so jank) to make because it uses so many of the tools and packages I developed personally. Instead of plain Remix actions, I used my own remix-endpoint for typesafe forms and mutations. Instead of Prisma, I used shitgen for my database. Instead of shadcn/ui (or whatever Vercel is using), I used my own design system. Honestly, using my own packages in common scenarios throughout my own projects has exposed loads of bugs that I wouldn't have otherwise found. For example, shitgen didn't have any support for enums, was improperly generating types, and failed to create certain queries involving limits and ordering. Not many people end up using my work because I have so little traction online, so it's important to me that I test my own work (with actual tests or in hands-on projects) to ensure their stability.

Challenges Faced

Finding a small model that fit my requirements was surprisingly difficult. I needed something

• small and fast enough to run on consumer hardware (my base Macbook Air)
• smarter than rocks just to follow what the prompt demands
• only output code, no text

Unfortunately, many LLMs include some alignment training and how they should respond to questions - they aren't designed to simply generate code. The smaller models at the time were clearly designed for autocompletions, and weren't capable of understanding my prompt. In hindsight, I probably should have used unsloth to fine-tune a model and force it to give me what I wanted with examples, rather than mucking about download and testing each one that I found on the Ollama directory. Ultimately, I settled for qwen2.5-coder:1.5b even though it's pretty bad at making user interfaces compared to ChatGPT. It's fast and kind of understand my prompts. It knows basic React and Tailwind, which was great for rendering a preview inside of an iframe.

Future Enhancements

Vercel's v0 has significantly improved and changed since I created my clone. I don't plan on keeping my copy up to date, especially when their team iterates so fast. However, I will probably look into implementing a projects tab, which is most likely a useful feature even for local and offline users. In the meantime, there's a lot to fix and implement. The real v0 supports selecting palettes and using images in prompts, which is currently not implemented in my clone. In fact, a lot of the UI elements present in my v0 clone aren't even functional because I dedicated most of my time into refining the generation and previewing features. I also want to experiment with selecting different models (something not available with Vercel's v0), which is especially important in a local environment when performance can differ wildly even within the same model size. Additionally, although I love Remix to pieces, I'm not sure about its future status now that it's been essentially integrated into React Router v7. I might have to completely overhaul the project and migrate it to the new framework mode of React Router, or I might just pick something entirely different like Tanstack Start (I'll wait for them to iron out all the bugs while it's in beta though). Honestly, I probably should have used Electron from the start...