Humind

Inspiration

I struggled with acne for years and tried countless products. As a man, my primary source of information was my friends, but what worked for them didn’t work for me. It's understandable because each person is unique and needs tailored assistance. I'm convinced that e-commerce needs a glow-up: an ultra-personalized, deeply human experience.

What it does

Humind mimics the in-store experience of a human consultant but online. On brand websites, you can FaceTime with a trusted consultant who shops with you, provides advice, and builds a crucial bond between the brand and consumer, which is essential in decision-making.

How we built it

I used Groq's super-fast inference with the Llama3 70B model for natural conversation, while Groq's function calling manages logic like recommendations and interface updates. The backend is built on Node.js + Express, with MongoDB for storing products and caching. The avatar was generated using Heygen.

Challenges we ran into

• Creating a real-time speaking avatar was challenging; after trying various options, I decided it would be impossible in 24 hours. Instead, I generated key video snippets and added the interface over the avatar while she talks. This creates the illusion of her being live.
• Summarizing YouTube videos was tough due to limited access to transcripts.
• Llama3 occasionally hallucinated and made up products and parameters, which I spent considerable time troubleshooting.
• Designing a compelling data model for the products was difficult, requiring multiple iterations.
• Fatigue led to unnecessary time debugging Groq streaming before realizing function calling is incompatible with streaming output.

Accomplishments that we're proud of

• The YouTube video summaries work surprisingly well and quickly with Groq, saving consumers hours of searching for reviews.
• The feature that retrieves products via photos is very effective.
• I'm overall proud of the product's UX/UI, finding it both functional and compelling.

What we learned

I learned a lot about open-source models, Groq, and the cosmetics industry. I had the opportunity to consult experts who taught me about RAG, fine-tuning, and function calling with LLMs.

What's next for Humind

I want to test the MVP with real users to gather feedback and validate its value. Exploring how voice-to-voice conversations differ from traditional e-commerce platforms will be interesting. I believe Humind can transform not only the cosmetics industry but e-commerce as a whole.

Built With

• angular.js
• express.js
• groq
• llama
• llama3
• mongodb
• node.js
• openai
• youtube