Dimitar Nedev

Data platforms

The platform is shared across Booking.com's business units. My part is mostly architecture, though I usually build the first working version with the team that will run it before handing it over. At the moment my work covers:

• Semantics. Semantic views, now in production, as the standard way into each analytical data product, so BI tools and AI agents use the same metric definitions.
• Access and governance. Attribute-based access control across the data warehouse and the lakehouse (row and column policies), and access that still holds when an AI agent acts for a person.
• Zero-copy integration. Governed data sharing with partners and across the group's brands, catalog federation, and zero-copy connectors to business systems. We built and tested these with the vendors, then rolled them out, and I co-wrote the reference architecture for sharing data with third parties. My bet is that within a couple of years most ETL and custom integration APIs give way to connectors and federation like these.
• Iceberg and catalogs. The Apache Iceberg lakehouse and how it works across engines, catalogs and storage, including Iceberg REST catalogs. I set the criteria for our choice of catalog.
• Serving. Getting analytical data to applications straight from the data warehouse at interactive speed, rather than from copies in separate stores.
• Finance-grade data. Auditability and controls for finance reporting, including SOx IT general controls and a zero-copy integration from finance systems into the data warehouse, now rolling out.
• Cost and reliability. Platform cost and query performance, failover, and privacy work such as deletion at scale.

We also work directly with vendors, including Snowflake, AWS and Immuta, on lakehouse interoperability and governance. That has included Open Semantic Interchange, an open standard for semantic models, which has since moved to Apache as Apache Ossie (incubating). The idea is that metric and dimension definitions can move between engines and BI tools.

I keep public notes on instructing coding agents and chat assistants, and on checking what they produce. At the core is a fairly large set of standing instructions that goes out with every conversation, and a way to check whether each change to it helps. The notes have a catalog of failure patterns, an audit prompt, a way to turn confirmed failures into regression checks, and sanitized copies of my setups for Claude, OpenAI, Cursor and Snowflake tools.

This is n=1 work, graded by me on the model versions I used, and it keeps the cases where the evidence went against my own instructions. The notes build on work at Booking.com and were released with its approval under CC BY 4.0.

There's an overview of the setup, the loop and what I've measured, and the notes themselves are on GitHub. The same page covers the instructions I use for data work and has results from a team benchmark of AI agents on Snowflake work.