The dashboard was wrong for six weeks and nobody noticed
A join quietly started dropping rows and the revenue dashboard kept looking plausible. What I changed so it never happens silently again.
Writing
Notes from the pipeline
Things I've learned building data platforms — what broke, what worked, and what the hype cycle gets wrong. New posts most months.
-
Aug 14, 2026
-
Aug 3, 2026
Semantic models are the most underrated work in analytics
Nobody puts "built a measure library" on a conference slide, but it took our reporting cycle from two days to four hours.
-
Jul 22, 2026
Iceberg won. Now what?
Everyone agrees on the table format now. That settles much less than people think.
-
Jul 9, 2026
What consolidating 8 source systems taught me about ELT design
Lessons from pulling product, marketing, and customer data into one Snowflake analytics layer without breaking anything downstream.
-
Jun 25, 2026
dbt tests are unit tests for your beliefs
You think order_id is unique. The test is how you find out when you stop being right.
-
Jun 12, 2026
DuckDB is my new scratchpad
Not everything needs the warehouse. A local engine that eats parquet has quietly changed how I debug and prototype.
-
May 28, 2026
Cutting our Snowflake bill without touching a single dashboard
Full-refresh models were rebuilding the world nightly. Making them incremental cut warehouse compute hard, and nobody upstairs noticed a thing. That was the point.
-
May 15, 2026
The anomaly detector that cried wolf
My first risk model was technically excellent and practically ignored. Fixing it had nothing to do with the model.
-
May 4, 2026
Data contracts: promising, mostly aspirational
The idea is right. The org chart is the hard part. Notes from trying a lightweight version in real life.
-
Apr 20, 2026
BigQuery optimization is mostly about not being clever
How warehouse workloads got ~45% faster with partition pruning, honest data types, and the removal of my own smart ideas.
-
Apr 8, 2026
Nobody reads your data docs. Write them anyway.
Documentation and lineage felt like chores until self-service went up 60% and my interrupt load collapsed. Turns out docs are read at exactly one moment.
-
Mar 25, 2026
LLMs in the pipeline: where they actually help
Ignore the text-to-SQL demos. The wins are smaller, weirder, and mostly about unstructured inputs and terrible source data.
-
Mar 12, 2026
From business analytics to data engineering: what actually transferred
I came in through the analytics door, not the CS door. Less of a handicap than the discourse suggests, with a couple of real gaps to close.
-
Feb 27, 2026
Kicking the tires on Microsoft Fabric
A Snowflake-and-dbt person spends real evenings with Fabric and comes back with mixed, specific feelings.
-
Feb 22, 2026
Boring pipelines are a feature
The most senior thing a data pipeline can do is nothing interesting. On idempotency, retries, and the luxury of not being needed at 3 a.m.