Glint: Reimagining E-Commerce

The worlds first 24/7 AI shopping channel! A realtime multimodal AI stylist that performs style session with users in real time, bringing personal style to live-streamed-ecommerce.

Comment

Inspiration

We really wanted to push the boundaries on what the next generation of shopping would be.

Looking at the trends in Asia we see live streaming as a massively growing market.

What it does

We've created an AI "celebrity" named Stella who performs real time styling sessions with the users. The content is created in a way which is also watchable by others,. with products showing on screen and being clickable. This allows us to create content which can be repurposed for both retailers and on a main shopping channel.

Our hackathon entry includes the core style session with each one recorded. The primary style channel will be a follow up to create with our winnings (if we are so lucky!).

How we built it

We used bolt.new to scaffold the base app. We then integrated a Tavus replica and gave her a prompt to become "Stella".

From there we self hosted our own LLM and integrated that via API.

The hard part thing was writing evals and end to end tests that verified we could return products and a transcript for Stella which allowed us to orchestrate the animation of products on the screen.

Challenges we ran into

Having the products transition on screen as Stella talks was quite tricky. It involved blending together the product overview stella returns with some in-app

Accomplishments that we're proud of

It feels like we created a person! One that can sell for us 24/7. Really strange feeling but we are super proud and we think the programmatic content generation aspect has huge potential to help smaller retailers sell product and create social media content,.

What we learned

How hard it is to distinguish between AI and humans now! Stella seems so real.

The giveaway is response latency.

The Tavus docs were a little confusing so we ended up hosting our LLM. This means the response time is a little slow. Reviewing it now I think we probably wanted to use their interruption API. We will try refactor that.

What's next for Glint AI Style Channel

So when we started out we figured it would be an "AI Stylist". That was before we seen how expensive Tavus is. So we think creating a style channel is perfect as it's basically programmatic content generation with a viral loop built in.

We aim to get the channel live and bring creators on board.

Built With

  • bolt
  • nextjs
  • supabase
  • tavus
Share this project:

Updates