How I used AI Video Generation for My Hackathon Submission

I used AI video generation tools Veo 3 Flow and Replicate to create a 3-minute hackathon demo video for NHS Buddy. Here’s what worked and what didn’t.

This blog post details my experience using AI video generation tools to create a compelling 3-minute video for my Kaggle - The Gemma 3n Challenge submission, focusing on a product called NHS Buddy. The product is an on-device medical assistant designed to support doctors and hospital staff with administrative tasks, critical alerts, and medical information, with the goal of “helping those who help us.”

Given the three-minute limit, it had to concisely explain the context of the product while also demonstrating its core features. Here’s the final video that we submitted.

The Video Storyboard: My Blueprint for Success

Before generating any video clips, I created a detailed storyboard. I broke each visual element into a “byte” or a single prompt. For example: a doctor checking a patient chart, receiving a critical alert, or reviewing medical data on a device.

This approach ensured that I had a clear vision for the final video and a systematic way to create the necessary clips that fit together seamlessly.

Veo 3 Flow: The Initial Go-To

My first choice for video generation was Veo 3 Flow. My experience with it was generally positive, and it helped me create a lot of the initial content for the video.

The Pros:

Prompt-to-Clip Efficiency: For the most part, I found that I could get the desired results within just a few attempts. I just needed to write the prompt, and Veo3 Flow would quickly generate a clip that matched my vision.

Batch Generation: A big plus was that it generated two videos at a time. This sped up the process, giving me more options to choose from and allowing me to progress through my storyboard more quickly.

High-Quality Output: The quality of the generated videos was impressive, which was crucial for a professional-looking submission.

The Cons:

Resource Intensive: The main drawback I encountered was the expense. AI video generation can be a costly process, and this was a significant factor in my decision to look for an alternative.

Inconsistent Outputs: While the model often retained elements I liked, it was a bit unpredictable. For example, if I wanted to introduce a new element in a subsequent generation, I would often find that the video had added the new element but removed something it was getting right all along. It felt like I was constantly trying to get the “perfect” video that had all the right elements.

Despite this, it was better at retaining elements than Replicate, which is why I often got the desired results within just a few attempts.

Replicate: The Backup Plan

When I started running out of credits on Veo3 Flow, I turned to Replicate. It was a more programmatic solution that offered access to the same Veo3 models, as well as others like the Wan model.

The Pros:

Cost-Effective: Replicate was a much more affordable option, which was a huge benefit, especially for a personal hackathon project.

Model Variety: It gave me access to a range of multiple models, which could be useful for a variety of projects and opened up creative options.

The Cons:

Programmatic and Less User-Friendly: The ease of use was a definite step down from the more streamlined interface of Veo 3 Flow. Since it was programmatic, it required more technical know-how and was a bit more cumbersome.

Inconsistent Generations (Again): The same issue I had with Veo 3 Flow persisted here, but it was more pronounced. I would often get two videos that each had a different element I needed, but neither had all of them combined. This meant a lot more trial and error to get the right clips

Slower Generation Time: It also seemed to be a slower process, likely because it only generated one video at a time, which reduced my overall output compared to Veo 3 Flow.

Key Takeaways

Both Veo 3 Flow and Replicate were instrumental in bringing my storyboard to life and producing a professional, engaging demo video for NHS Buddy. The main lessons I learned:

• A storyboard is essential.

Having a clear plan made the generation process much smoother.

• Expect iteration.

AI video tools are powerful but unpredictable. Getting the right clip often required trial and error.

• Balance quality and cost.

Veo 3 Flow delivered higher efficiency, while Replicate was more budget-friendly. Using both strategically worked best.

Ultimately, these tools allowed me to build a clear, compelling narrative within the hackathon’s three-minute video limit.

As AI video generation continues to evolve, I see it becoming a standard tool for hackathons, prototypes, and product storytelling.