Finding leads in the noise of social media

Collects public posts and ranks them with embeddings and a score, delivering only likely customers along with the reason.

Context

Founders and sales teams lose hours looking for people asking for exactly what they sell, scattered across thousands of posts. Sinalo collects public conversations and delivers only the ones that look like a lead, explaining why.

Constraints

Almost everything is noise: the share of useful posts is small.
Sources with different formats and limits.
Each LLM classification costs money; volume has to be filtered first.

Decisions and trade-offs

Embeddings before the LLM. A first semantic-similarity pass discards cheap noise and leaves the more expensive model only for candidates.
A score with a rationale. Users get a score and the reason, not just yes or no. The explanation builds trust and helps tune the criteria.
A staged pipeline. Collection, cleaning, embeddings and scoring are independent, so a source or a model can change without redoing the rest.

Result

Product launching, with a public landing page at osinalo.com.

What I’d do differently

Build a hand-labeled evaluation set from the start to measure score precision on every model or prompt change.