Member-only story
Engineering Multi-Agent AI for a 100 GB Data Warehouse
Tackling accuracy, security, and speed at scale
Read this story for free: link
Enterprises now let business users question their data warehouses through AI agents. At warehouse scale such an agent carries four requirements at once: accuracy that a finance team can sign off on, strict governance over the data it may touch, fast access to hundreds of millions of rows, and response times people will accept. A typical agent falls short on all four as the data grows. In this post we build one on 100 GB of retail data, one measured fix at a time:
To test this architecture properly, we have to replicate it on consumer hardware first. I am going to use the most minimal setup I can: one 16-core laptop with no GPU, a single DuckDB file and Qwen models behind an OpenAI-compatible endpoint. These are the pieces we build on it, in the order we need them:
- A warehouse on a laptop: 959,031,513 rows in one 28.6 GB DuckDB file.
- A ruler: an answer comparator and an exam with trusted answers.
- Three walls: a statement guard with its own row estimator, a locked connection and runtime limits.
- A model client: a cache, retries and a ledger in credits.
