Skip to content

// the tools behind the work

Uses

The tools I actually use to build data systems and the agents on top of them. Most of the agent pieces are my own, and open source.

Coding agent

  • Claude Code is where most of my work happens: it reads the repo, runs the commands, and I review every change before it lands.
  • Skills I wrote for repeat jobs: wrap-it-up keeps a per-repo memory so the next session starts without a recap, and html-slide-deck turns findings into a single-file deck.
  • MCP servers that give the agent read-only access to the data I check every day, so a quick question doesn’t need a dashboard.

Guardrails for agents

An agent with a terminal can spend money and leak secrets as fast as it can type. These hooks make it stop and ask.

  • warehouse-brake: asks before a command can spend Databricks or Snowflake credits, including through dbt.
  • az-brake: asks before an Azure CLI command creates or starts something that bills.
  • commit-guard: blocks a commit when the staged diff looks like it holds a real secret.
  • A pre-commit hook that rejects any file over 5,000 tokens, so every file stays small enough for an agent (and a person) to read whole.

Small things that make agent work nicer

Data

  • Lakes and engines: Apache Iceberg, Trino, StarRocks, DuckDB, Parquet.
  • Platforms: Databricks and Snowflake, deployed as code.
  • Pipelines: Apache Spark, Kafka, Delta Lake, dbt, and pyspark-quarantine for bad rows.
  • Databases: PostgreSQL first, Redis when something needs to be fast.

This is a uses page. See also what I’m working on now.