Sentimeter

Inspiration

Everyone talks about AI these days on Twitter and it's impossible to understand the underlying trends & phenomena in the moment, especially with all the noise, smoke & mirrors, and manufactured narratives. We wanted to look from a birds eye view to unravel these patterns, and to understand why and how sentiments spread online.

What it does

User inputs a tracking objective (eg. How do people perceive AI companies, specifically OpenAI, Anthropic, xAI, etc.), and Sentimeter fetches live, relevant Twitter/Bluesky posts and assigns it a sentiment score, weighted by the post's engagement. It then visualizes the data (sentiment across time, standard deviation). Users can query specific parts of the graphs, and an agent answers by querying granular post data on our DB, either by text or voice.

How we built it

We have a real-time post processing pipeline (large-scale text filtering, assigning sentiments, classifying by subtopic), a data visualization UI, and a agent harness that has access to everything the user can see.

Challenges we ran into

Being able to support vast stream of data in real time & cheaply. For a simple question like "Sentiment on AI companies" we have process over 800k posts, costing 1.5B tokens.