Hey, I have been coding and preparing for the ATmosphereConf until the last second, so I did not have the time to write a blog post.
Please just check out divepool.social, the divepool feed on Bluesky or contact me with any kind of inquiry/feedback. Deeply appreciated!

You can also read the following slides to get some feeling for the tech I am using or the nerd I am.

Disclaimers

• My level: ELI12 (lateral mover)
• Respecting !no-unauthenticated
• No LLM training (yet)

TL;DR - the stack

TL;DR - the stack

Like(d) about Bluesky (Twitter)

1. "Do we match?" (platonically) answered in seconds.
2. Quantity over quality
3. Long tail discovery
4. Lateral navigation
5. Data access

Follower graph

TF-IDF

Keyword extraction by comparing term frequencies in a document with those across all documents

TF-IDF

Highly individual tokenizing pipeline (is that normal?)

• Language detection (about 3% Bluesky mismatch rate)
• Cleaning, hashtag splitting, stemming, ngrams…

Embeddings

• Magic
• Embed all the posts (that are English or German and longer than 120 chars)!

Naive approach

You are one (average of all your posts).

More nuanced approach (used by divepool discovery feed)

You contain multitudes (of post clusters).

How divepool.social works right now

Start from a random spot in the embedding space, retrieve the closest 1000 posts (1 per account), cluster them, show medoid per cluster, repeat.

Topic labels?

-> TF-IDF on embedding clusters

Slightly better labels?

-> Embed and rerank terms

Time for demo?!

Actually useful labels?

Outlook: fine tuned labeler?

(Smallest Qwen 3.5 already only takes 1 second for labeling.)

Further outlook

1. Accessibility
   • improve UX
   • native apps (Expo?!)
   • API access?!
2. Performance
   • embeddings are instant, indices are not (yet)
   • more stable setup (backups!!!)
3. Business
   • is there a (B2C) product somewhere in here?