eye← back
pipeline

how eye sees

EYE is a realtime attention market. To price attention, we first have to measure it. Every minute, EYE pulls fresh signals from the open internet, normalizes them, scores them by velocity, and dedupes the noise. The output is a single ordered feed of what the network is actually paying attention to - and that feed powers everything: the watching wall, market detection, and ranked narratives.

01 - ingest

Every 60 seconds, a scheduled job hits 18 public endpoints in parallel - feeds the platforms publish themselves.

X / Twittersocial

real tweets via rotating Nitter mirrors, search filtered by min_faves

Farcastersocial

channel feeds via Neynar (ai, crypto, base, founders…)

Redditsocial

hot posts from narrative-rich subs

Hacker Newstech

top stories, ranked by score + comments

Lobsterstech

hottest stories with weighted comments

Dev.totech

top articles by reactions

Product Hunttech

daily launches via RSS

GitHub Trendingtech

what builders are starring right now

arXiv (cs.AI / LG / CL)research

fresh ML papers, last 24h

CoinGeckomarkets

trending coins with 24h price change

DefiLlamamarkets

biggest 1d TVL movers above $1M

DexScreenermarkets

trending pairs, volume-weighted

Polymarketmarkets

active markets ranked by 24h volume

TechCrunchpress

RSS firehose

The Vergepress

RSS firehose

Decryptpress

crypto press RSS

02 - score

Each signal gets a velocity in [0, 1]. The exact formula varies by source - likes + 2×reposts + replies for social, log-scaled volume for markets, abs(% change) for price movers - but the shape is the same: normalized engagement relative to what's typical for that source. Velocity is what lets us compare a viral tweet, a 40% TVL spike, and a top HN post on the same axis.

03 - fuzzy dedupe

The same story shows up 6 times in an hour: a tweet, a Reddit thread, a TechCrunch headline, three other tweets quoting the first. If we kept all of them, the feed would feel like spam and the rankings would be gamed by sheer repetition.

Before any signal is written to the database, we compute its 3-gram word shingles, then measure Jaccard similarity against every signal captured in the last 12 hours. Anything with similarity ≥ 0.7 is dropped as a near-duplicate. The original wins; the echoes get suppressed.

04 - rank

What survives dedup gets upserted into a single signals table, trimmed to the top 1500 by velocity. The watching page reads it live. The rankings page groups by source and sorts by velocity. The markets page detects clusters of related signals - that's where narratives become tradable.

why this powers eye

A narrative isn't a tweet. It's the moment a tweet, a chart, a commit, and a headline all start saying the same thing within a few minutes of each other. You can't see that by reading any one feed. You see it by watching all of them at once, weighted by velocity, with the duplicates collapsed.

That's the input EYE prices. Without clean ingest, every market on this site would just be a popularity contest for whoever shitposts loudest. With it, attention becomes legible - and what's legible can be measured, ranked, and eventually traded.