Notes

The obvious answer to "institutional-grade on-chain data" is to buy it. Providers charge $15K–30K/month for what we generate ourselves on under $100/month. The difference isn't magic — it's a caching layer that respects rate limits.

We ingest 17 blockchains and score 1,000+ coins daily. The naive approach is to hammer every RPC endpoint every few minutes. That blows past free-tier limits immediately and turns the cloud bill into a second mortgage.

Instead, every fetch goes through a cache that batches reads per endpoint, respects each chain's rate window, and marks freshness so the scoring jobs only re-pull what actually changed. Partitions on BigQuery + a single lean Cloud Run instance keep the compute honest. The result: the same data institutional teams pay a fortune for, fresh inside a 5-minute window.

A scoring model is only as good as its priors. We settled on 40/30/30 after watching what actually moves markets over months of live data, not from a textbook.

OnChain (40%) leads because it's the hardest to fake. Exchange balances, whale moves, network activity — this is behavior, not narrative. It's also the least-covered signal in retail tools, so it's where the edge lives.

Value (30%) grounds the score in fundamentals — what the network actually produces relative to its price. Momentum (30%) captures what the market is doing right now. Momentum gets the smallest slice on purpose: it's the noisiest signal and the easiest to get burned by.

These weights aren't sacred. They get re-validated against live outcomes as part of the evaluation pipeline — but the ordering has been remarkably stable: fundamentals first, market sentiment last.

An LLM asked to "map my emotions" will cheerfully invent them. For a journaling product, that's not a bug — it's a betrayal. So Pratyaksha runs 4 agents that keep each other honest.

The first agent extracts what happened. The second infers emotional state. The third looks for recurring themes across entries. The fourth is the skeptic — it cross-checks the others' claims against the raw journal text and rejects anything it can't find evidence for.

That final validation layer is the whole product. An emotion map the user can't recognize in their own words is worse than no insight at all. By forcing every output through a grounded check, the pipeline produces insights that feel earned — because they are.