We built "Vibe Draw" at GenAI Genesis 2025 and won Best Use of Gemini. It's basically a web app (supports PWA on mobile devices + joystick controls) that turns your sketches into detailed 3D models, which can also be exported into .gltf files. You can refine your drawings in an infinite 2D canvas, edit existing 3D stuff, and mess around with the scenes all in one place.
I was unwillingly tasked to do all the 3D part, since no one (including claude 3.7 sonnet) couldn't do it. It took me 4-5 hours alone trying to make raycasting work properly with individual group and mesh selection, custom TransformControls, and a ton of other mesh material editing stuff. But the best part of my experience is when I finally got to test Cerebras Systems in production. It was generating 3800-4000 tokens/s for code extraction and concatenation and completely blew my mind. This is the future of autoregressive models, and it's brighter than ever.
The wildest part of this whole thing was working with Osher Ahn-Clifford. One random day, Martin and I were discussing who to work with for GenAI Genesis. Then he told me that he heard rumours of Osher asking to attend the event, and reached out to him on LinkedIn asking if he wanted to team. Being the cold reachout goat that he is, Martin successfully scheduled a meeting with him and his parents. Before we knew it, Osher was part of our team. For context, Osher is 11 years old and is currently studying 2nd year Computer Science at University of Waterloo. He came to the hackathon on day 2, and instantly caught everyone's attention. Our app idea was there but it was pretty barebone at the time, having almost no 3D features implemented. After we explained the codebase to him, we assigned him some basic tasks like adding buttons, updating some states, etc. He was really good at CSS and adapting to the code, but his level of focus and attention to detail was what surprised us the most. He could notice 1-2px offsets, had a good sense of design, and implemented the features we asked him to effortlessly. There are some moments during the hackathon when I genuinely forgot that he was 11, since he's quite literally more qualified than some Master's students I've met at hackathons. But Osher is only human, afterall. Since he didn't have a lot of experience with backend (reminder that he is 11), Ahtesham had to do most of the backend. Our backend was insanely complex, Martin and Ahtesham designed this architecture with Redis, SSE, docker compose, celery, and FastAPI. We had task-scheduling systems for the image generation, 3D creation, model editing, etc. I only briefly worked on the Gemini pipeline and implemented the cerebras endpoint, but making the backend felt very nice.
I also tested Microsoft's open source 3D model called trellis and it was insane at generating 1:1 replicas of 3D models from 2D images. But Claude didn't disappoint either, the one-shot capabilities of 3.7 sonnet is actually beyond human comprehension and it makes me want to work for Anthropic more and more (please hire me). Gemini 2.0 Flash with its image generating and editing capabilities was also insane, it is the fastest imagen I have ever used. Google Deepmind better hire me soon, I want to work on this so badly.
Of course, because the universe has a sense of humor, our CDN decided to die during our second demo. I was so confused during the demo why the iframe did not load despite having 0 error logs, but the local version of three.js was somehow working like magic. I discovered after that demo that the esm.sh build API (which was hosting the three.js build for our iframes) had a 1 hour downtime (99.93% uptime for the last 3 months btw). Very unlucky, but we still won so I'm not complaining.
Now, let me expose GenAI Genesis organizing team. The organization was worse than GeeseHacks. The judging rounds went fine, but that didn't matter since the organizers were literally fighting over the slides, editing them every 10 seconds. I heard rumours that some organizers completely ignored the judging data, and argued that "my friend's project should win!11!!!1". Nepotism might have won at the end, since the first place project was one of the worst LLM wrappers I've ever seen.
3 out of 4 members on the first place team were organizers. Yup, you heard it right. The FIRST PLACE WINNERS of GenAI Genesis consisted of 75% organizers.
Organizing clubs:
Person 1's LinkedIn:
Person 2's LinkedIn:
Person 3's LinkedIn:
But since we won anyways, I guess it's just a matter of skill.
Hackathon organization 3/10, actual experience 10/10.