Hi — I’m Kai

For onchain analysis, the limiting factor is often SQL and data plumbing. Researchers end up spending too much time writing queries and moving data between Dune and external APIs. At Blocmates, we solved this with an agent-driven ETL pipeline that automates dashboard production for the research team. You fill out a data spec, the agent writes, tests, and pushes queries to Dune using GitHub for version control. If the spec involves heavy API usage, the agent sets up a GitHub Action to collect data and load it into Dune. Queries are generated in minutes and are then ready for visualization in Dune’s UI. ...

Since the early days of GMX, the market has treated the token as the equity-like asset and the LP as the yield product. The token captures the narrative and gives investors a liquid claim on protocol growth. Meanwhile, the LP is usually evaluated on its APY. That framing misses the risk structure of the exchange. Every leveraged position creates exposure that has to be priced and resolved, and in pool-based designs, the LP carries that exposure. When the system prices risk well, trading revenue accrues to LP NAV. When it does not, trader P&L from toxic flow hits the LP first. LP NAV therefore becomes a live mark on the protocol’s underwriting quality. ...

With Crypto operating 24/7 it’s challenging to keep up with current news while also going deep on a report. Writing forces you to focus on a project or sector deeply, but you still have to track headlines that could change your thesis or reframe the data. To solve this, I vibecoded an AI research assistant to keep me up to date with the latest news in crypto and finances everyday, in around 20 minutes reading time – down from 1-2 hours. It runs on locally hosted models including OpenAI’s GPT OSS 20b, and Deepseek R1 via LM studio. ...

@aave has grown to become DeFi’s Trojan Horse, reaching a scale comparable to traditional banks. With $35 billion in TVL and $57 billion in net deposits, it now controls 25% of DeFi TVL and ranks among the 50 largest US banks by assets under management (AUM).

While building the MetaDAO dashboard at Blocmates, @563defi and I spent some time digging into MetaDAO’s launchpad program to understand the existing Dune queries, build on them, and create new ones. These are my notes, which I will be referring back to as I verify & build out more queries on MetaDAO’s activity. This can also be used as context to feed to an LLM for generating simple queries. Hopefully this helps you understand the program structure and verify your queries are pulling the right data. ...

Frontier model training has consolidated into a small set of operators with the capital to assemble tightly-coupled GPU clusters. That concentration produces the fastest path to capability, but it also limits the decisions about what gets trained, on what data, and on what terms to those same operators. For workloads that prioritize censorship-resistance or sovereign infrastructure, teams require frontier-level models that are structurally open. However, existing open-weight releases like LLama or Mistral are only open at the discretion of the sponsor. ...