RocketScan.live

The problem

Anything built on live social data hits the same three problems: the crowd is often wrong, some of it is deliberately coordinated to be wrong, and reading enough of it to find out is metered by the post. Most demos ignore all three and present a confident answer built on whatever the firehose happened to return.

Our approach

We picked a domain with a hard, public ground truth - a rocket either launches in its window or it does not - so the system could be graded rather than admired. RocketScan fuses live chatter, on-site pad photos, weather, and schedule feeds into a GO, HOLDING, or SCRUB LIKELY call ahead of the official announcement, then scores itself against what actually happened.

Credibility weighting, not vote counting

Sources are scored, so one verified observer at the pad outweighs a swarm of accounts repeating each other.

Coordinated manipulation is detected and discounted

A deterministic anomaly detector flags brigading and verbatim repost swarms, and the verdict's confidence drops instead of being amplified by them.

Every paid call is metered and capped

Spend is itemized per analysis, reads are launch-scoped and incremental, and a shared cache means cost scales with time and distinct launches rather than with visitor count.

Graded against real outcomes

Hit rate and median lead time before the official call are tracked, so the system's accuracy is a published number rather than a claim.

The whole product demos with zero keys

Every paid call has a mock path and the UI labels its sources honestly - live or fixture - so nothing is ever silently presented as real data when it is not.