Into the Brand Universe

Inspiration

I run a business that operates in fintech/ AI, typically mining regulatory documents, and deriving insights from disparate data sources. I wanted to explore what kind of insights twitter data may be able to give a brand on their trends or their customer segments.

What it does

Visualize, cluster the twitter universe for a brand/ topic.

How we built it

NextJS, Express JS, Postgres Firebase for Auth UMAP for clustering Qwen3 8B for embedding (nice since it offers 32 dimensions) XAPI for users, posts (also has streaming capabilities, but less interesting for brands)

https://www.loom.com/share/6906fcbef0404ec19e8b4e02324b7e78

Challenges we ran into

Persisting back projections (to 3-space) at the right frequency Making embeddings useful Using the appropriate dim reduction algorithm

Accomplishments that we're proud of

Satisfied that we made a navigable cloud that allows us to explore intuitively the world for a brand.

What we learned

Insights for twitter are VERY often polarized. Even seemingly loved characters like Jackie Chan are dual clustered along love-hate lines

What's next for Into the Brand Universe

AI analysis of clusters Top on the priority list would be "What-ifs" using sampling methods. What if as a brand I posted XYZ tweet, how would the clouds move/respond. Can be simulated with LLM response