Pixie

• 

  Lab View (Canvas) - Desktop

Inspiration

Only 10 million drugs have been synthesised and tested - ever.

Most molecules are yet to be experimented with. This makes them both novel, unique and interesting. We propose a systematic way to go about the discovery and research of molecules so that it's both fun, consumer friendly and unlocks a means of shared resource.

What it does

Users get an open canvas (their "lab") where they can randomly add known molecules or use AI to generate new ones at random. Every molecule is researched and summarised with interesting facts using Brave and Mistral prompt chaining - plus a unique Stable Diffusion generated asset.

Users can combine molecules to synthesise new ones, view them, then decide if they want to save them to their lab.

Users can chat to Pixie to learn more about their entire lab using "chem chat".

In a world where LLMs can be run locally with comfortable confidence and results (think using your browser via agentic workflows) we can have our user base learn, have fun, collect assets and be contributing to the over 1 duotrigintillion drugs to research!

How we built it

• Mistral for molecule generation
• Stable Diffusion on Nebius for image/video generation
• Brave Search API for molecule research
• SMILEs Data Set
• SvelteKit web application
• PartyKit for real-time collaboration

Challenges we ran into

• Functional PartyKit collaboration in production
• Consistent image generation (deterministic)
• Sleeping
• Combining Brave/Mistral to prevent verbose outputs (yapping).

What's next for Pixie

• Share the project, track users, track feedback.
• Document all molecules discovered and their findings.
• Build AI agent that systematically considers findings and instead of random generation determines interesting molecules to generate... so a real drug discovery journey.

Built With

• brave
• mistral
• nebius
• node.js
• partykit
• stable-diffusion
• sveltekit