On-chain data engineer. I build pipelines on EVM chain data in Python and SQL, and I check what they produce against the chain itself before I trust a number.
A lot of my work so far has been about the errors that don't crash anything: an explorer API that answers HTTP 200 with an error inside, a page that stops at 1,000 rows, a missing price read as 0, a rate limit reported as missing data.
Findings page with charts and the data behind them: am0muk.github.io
onchain-tieout: rebuilds a wallet's ETH and ERC-20 balances from Etherscan V2 history and checks them against JSON-RPC state at the same block, with a probable cause for every difference. On a public wallet (vitalik.eth) it produced 10,476 balance rows, 8,281 of them exact. Its first live run found bugs in its own code that the offline tests had missed: pages capped at 1,000 rows, a block too large to paginate, and rate-limit errors reported as data.
mev-scout: read-only measurement of MEV on EVM chains. Aave V3 liquidations over 12 months on 12 chains, same-chain DEX arbitrage on Arbitrum, and cross-chain arbitrage between Arbitrum, Base and Optimism. The pass/fail rule for "should a newcomer build a bot in 2026?" was fixed before any data was collected. No niche passed.
lp-sim: simulator for Uniswap V3 range strategies. It rebuilds per-tick liquidity from Mint and Burn events, and its fee accounting matched the pools' on-chain fee counters to within 0.3% on three stable pools, the largest with 62,484 swaps. 1,464 strategy runs on 8 stable pools, none passed the 15% target fixed in advance.
healthfactor_watch_bot: a read-only Telegram bot for people who borrow on Aave. It alerts when a health factor falls to levels the user sets, on Aave V3 (20 chains) and Aave V4 (4 chains), with opt-in stablecoin depeg alerts. 230 automated tests, MIT license, runs in Docker. You can try it at @healthfactor_watch_bot. Review caught three defects before release: a price source that answers HTTP 200 with an empty result for an unknown token, a freshness limit that silently dropped 5 of 12 stablecoins, and an alert that was lost when sending to Telegram failed.
A private multi-chain pipeline (16+ chains, 20+ protocols, 1,700+ tests) runs a balance check against the chain after every run. I used it for my own German tax return.
I write the specification and the test plan. AI coding agents write much of the first version. I review the code and check the results against real chain data, and I publish the verdict even when the answer is no.
Python · SQL (PostgreSQL, SQLite) · Dune · Uniswap V3 math · Etherscan and Blockscout APIs · archive RPC · Telegram Bot API · Docker · Grafana · Linux
Articles on silent failures in DeFi data (LinkedIn, DEV.to, Paragraph).