Cloud-Clips Automator

Inspiration

Cloud-Clips, an early-stage thought leadership services startup I founded, aspires to provide semi-automated how-to video creation services to leading cloud providers like AWS. By streamlining creation of how-to videos through automation, we aim to make it easy for AWS customers to access demos of the latest and greatest AWS features as soon as they are released.

I publish how-to demos regularly on our dedicated YouTube channel: https://www.youtube.com/@cloud-clips-by-Kanchan. To build these videos, I leverage our homegrown secret sauce - the Cloud-Clips engine, which operates on markdown instructions and automates a significant portion of the video creation process.

Necessity - The Cloud-Clips engine requires video instructions in a specific format, which differs from the AWS markdown format. Until now, I have had to manually convert steps instead of directly reusing steps from AWS documentation. This process is time-consuming.

What it does

The PartyRock project automatically converts AWS documentation markdown into Cloud-Clips engine compatible markdown. This eliminates the need for manual input, allowing us to leverage technology for building Cloud-Clips compatible instructions.

Project also writes SEO optimized titles, video description, and identifies hashtags for YouTube.

How we built it

Building PartyRock was simplified by the fact that PartyRock is a free service. While experimenting with Bedrock, I realized that Titan models are updated regularly, making them a natural choice as foundation models for multiple functionalities. Claude performed exceptionally well on most tasks assigned. To ensure good, consistent, and repeatable results, I set the top_p parameter to 0.9 for all widgets.

I invested significant effort in the AWS service extraction step. Despite exploring various models and prompt engineering techniques, I ultimately needed to provide an exhaustive list of AWS services to achieve reliable results.

Overall, the process was a valuable learning experience, providing insights into the behaviors of different models and the effectiveness of prompt engineering techniques.

Challenges I ran into

AWS Service Category Identification: Most models struggled to identify the AWS category to which an AWS service belongs. They even lacked awareness of the existence of Amazon Bedrock. Llama, Claude, and Jurassic incorrectly reported that a recently released AWS service did not exist.

Markdown Conversion: Converting simple markdown to Cloud-Clips compatible markdown using one-shot/few-shot learning techniques proved challenging. Production use would likely require fine-tuning to ensure the underlying foundation model understands and applies all the various markdown customizations.

Accomplishments that I am proud of

The results were impressive! I tested the app with various inputs, and it performed very well overall. As shown in the demo, I was able to get model return an output which I could directly feed to the cloud-clips engine.

What I learned

I gained a tremendous amount of knowledge about prompt engineering, experimenting with Bedrock models at no cost, understanding the capabilities of different models, and refining responses using various techniques. Additionally, PartyRock's free tier allows us to continue leveraging its benefits.

What's next for Cloud-Clips Assistant

We would love to finetune and use bedrock models in production for converting simple documentation into cloud-clips compatible markdown format.

Built With

• bedrock