Significant Speed Difference Between HeyGen Site Avatars and My Implementation

Hello,

I'm currently running a project that uses HeyGen's interactive avatar technology. I've implemented it exactly as described in the official documentation and deployed it on a Vercel server.

Recently, I tested some of the new avatars directly on the HeyGen website (using the same training data as my implementation) and noticed something striking: the avatars on your site respond significantly faster than my deployed version.

On my Vercel implementation, there's an unbearable latency when interacting with the avatar, while the same interactions on your website feel much more responsive and fluid.

I'm wondering:

Are there any server-side optimizations or configurations that I might be missing in my implementation?
Have there been recent performance improvements to the avatars that aren't yet reflected in the documentation?
Is the difference in speed possibly related to hosting on Vercel versus your native environment?

I'd appreciate any insights on how I might improve the response time of my implementation to match the excellent performance I experienced on your site.

Thank you!