Shadow Social

Shadow is an intelligent assistant that manages your entire social media presence. It keeps your channels active, relevant, and consistently high-quality, handling everything, so you don't have to.

Comment

Inspiration

I ws inspired to create Shadow after seeing lots of friends with small businesses struggle to maintain a quality and current social media presence. Most smaller businesses don't have the budget, nor the time to constantly maintain multiple social media channels. They also do not have time or budget to execute photoshoots and create content on a regular basis. This is where Shadow steps in.

What it does

Shadow is an intelligent assistant that manages your entire social media presence. It keeps your channels active, relevant, and consistently high-quality, handling everything in the background, so you don't have to. It connects to all your social networks. You choose your brand, product, theme or trend and set up as many virtual 'influencers' as you need.

Then, with one click you can trust Shadow to run your campaigns or social media posts. All you need to do is check on your phone and review before scheduling.

How we built it

I have decades of experience as a senior developer, working on Macromedia Flash Server and on prestigious campaigns and platforms for some of the largest brands in the world. I used a combination of LLMs to assist with coding, in particular Qwen. I was impressed by the range of generational modela available such as WAN2.7 & Happy Horse and have tried my absolute best to ensure content generated passes through pipelines which audit and self-heal content that has been generated by the models. This ensures that users do not encounter hallucination any where close to regular levels.

We are using PostGresQL, SQLLite ReactJS, Qwen, AlibabaCloud and Cuda to run the platform.

Challenges we ran into

One of the major problems was character drift & AI hallucination using WAN2.7 and any major LLMs. I have spent the majority of the project trying to make humans as human as possible, and videos as natural feeling and as closely linked to the chosen brand/theme/product/trend as we can, given the limitations of models. In order to do this I built a NVIDIA GPU instance which is able to run face swapping libraries, audio models (except for Cosy Voice which we use) and ffmepg for video stitching.

The hardest part was lip syncing, character drift, battling with the token limits to prompt as best as possible and using the image arrays where models allowed.

Accomplishments that we're proud of

I am really proud that i have managed to pull this together in a short time, but harnessing my decades as as software engineer and senior creative, along with tools like Qwen Code I have been able to build this in record time.

What we learned

I learned that LLMs have a way to go yet before they are perfect. There is still a huge problem with hallucination and accuracy, especially with generational models.

What's next for Shadow Social

Its ready to scale and is performing well on AlibabaCloud. Next for me is to market the product, expand the user base and improve the accuracy of generated assets (images, video and copy). I also want to work on the agentic side, bringing in auto-post commentary & seeking modes where the influencers actively target specified markets for users.

Built With

Share this project:

Updates