NightWatch
A Graph subgraph for Blur's Blend protocol, paired with an AI skill for loan portfolio analytics.
Demo
About
This project indexes Blur's Blend NFT-lending protocol on The Graph, then layers a Claude Code skill on top that turns raw on-chain loan data into real portfolio analytics no explorer gives you out of the box. The subgraph tracks a Blend loan's full lifecycle from a public GraphQL API: origination (LoanOfferTaken), repayment, refinancing, Dutch-auction exits, seizure, and buy-to-borrow-locked sales — fully synced against production mainnet history, not a sample window. The AI skill queries that data and computes what actually matters to a lender: APY (validated against Blend's real contract source rather than assumed), USD-converted gains to date plus monthly/annual forecasts using the protocol's actual continuous-compounding interest formula, LTV via live NFT floor prices (looked up once per collection, not per loan), and auction status with an accurate countdown.
How it's made
Stack: The Graph (AssemblyScript mappings + GraphQL schema, deployed to Subgraph Studio), Solidity (Blend's own verified contract source, pulled from Blockscout), Matchstick for unit tests, and a Claude Code skill (a Markdown instruction file, not custom backend code) as the analytics layer on top. The subgraph has two data sources on the same proxy address — Blend's business-logic ABI plus the ERC1967 proxy's admin/upgrade events (deliberately no-op handlers, confirmed dormant against real logs) — feeding a schema modeled to mirror the contract's own storage pattern: one mutable entity per lienId, overwritten across its lifecycle, the same way the contract itself only ever stores a hash per lienId rather than a history. The skill itself is fully keyless at runtime: CoinGecko for floor price/ETH-USD, a public RPC for on-chain name() fallback and current block height, all stitched together with curl and short Python snippets rather than a full web3 SDK — including hand-decoding ABI-encoded eth call results byte-by-byte. We iterated fast by temporarily narrowing the subgraph's startBlock to a ~3-day window during development — full resyncs took minutes instead of the multi-hour cost of the real ~8.76M-block production history — then restored the real startBlock once the schema and queries were verified correct.