AI-graded signals, in real time.
A real-time crypto market dashboard where every signal — price spike, on-chain move, social burst — gets graded and explained by a Claude reasoning pass before it ever lands on a trader's screen.
The problem
Crypto traders live in a fragmented stack. Coingecko for price, Glassnode for on-chain, Lunarcrush for social, Discord and Telegram for narrative, plus three different alert systems firing into their phone. The signals exist. The synthesis doesn't. By the time a trader has triangulated a move across four tabs, the move is over.
The opportunity wasn't another dashboard or another alert system — it was a layer that watched all of them and gave the trader one ranked, explained read of what was actually happening.
What we built
Blockview ingests every market's price tick, on-chain transfer above a configured threshold, and social mention burst into TimescaleDB at sub-second resolution. A Claude Sonnet 4.6 grading pass runs over every candidate signal — scoring novelty against the rolling history, explaining the contributing factors in one sentence, and ranking it against everything else currently active. Only the top-N graded signals stream to the live dashboard. Everything else stays queryable but quiet.
The dashboard itself is a single screen: a heatmap of active signals across markets, a streaming feed of the top reads with Claude's one-line explanation attached to each, and a pinning workflow so a trader can save a thesis and have the agent re-grade it against new data without re-querying anything. Portfolio risk view layers on top — same signal grader, scoped to the assets you actually hold.
Before Blockview, I was watching four tabs and missing the move. Now one screen tells me what to look at, why, and how confident the model is. The rest just falls into the background.
How it shipped
Three weeks of feed plumbing — pulling price, on-chain, and social into one TimescaleDB schema with a unified event shape. Three weeks of grader development — calibrating the Claude pass against a backtest of historical "this mattered" events tagged by an in-house analyst, then tuning the prompt and the cost-per-grade math until the streaming budget was sustainable. Two weeks of UI. Three weeks of closed beta with thirty traders providing daily feedback on every Claude grade — that loop is what got the false-positive rate down to where streaming the top-N became viable.
Pricing is a flat monthly per-seat plus a metered fee on portfolio risk re-grading. The infrastructure cost is largely the grading inference — keeping that line predictable was the engineering work that mattered most after launch.