Overview & result
Firehose Judge samples the live Bluesky stream and asks Jev eight typed questions per post in one call. A Cloudflare Durable Object filters unsafe posts and sends uncertain judgments to a visible human-review lane; the author reports about $0.00003 per post.
From the original post
Put typed, calibrated judgment on a live stream. This repo watches the Bluesky firehose and asks Jev eight questions about every sampled post in a single call: intent, topic, sentiment, engagement bait, hostility, sarcasm, bot-ness, and whether it is safe to put on a screen. The answers come back as typed distributions with confidence, in a few hundred milliseconds, for about three thousandths of a cent per post.
How Jev fits in the loop
- Sample incoming Bluesky Jetstream posts in a Cloudflare Durable Object.
- Ask Jev eight typed questions about intent, topic, sentiment, bait, hostility, sarcasm, bot-likeness, and screen safety.
- Drop posts judged unsafe before broadcasting to browser tabs.
- Show lower-confidence judgments in a lane marked for human review.
How to reproduce
- Connect to the Bluesky firehose websocket stream using the project setup.
- Configure the eight typed questions in questions.json and a local TypeSafe key.
- Run the Cloudflare Worker and observe the confidence distributions, unsafe-post filter, and human-review lane.
Why this build matters
Shows how System 1 decision engines make real-time event-driven agent architectures economically viable.
Reported performance
Reported by authorCost: ~$0.00003 per post
Limitations
- The project samples posts with a token bucket; it does not classify every post in the public firehose.
- Classification and cost figures are author-reported, not independently reproduced.