Posts
All the articles I've posted.
-
dbt + DuckDB: the good, the bad, and the workarounds
Running dbt 1.11 with DuckDB 1.10 on a 167k-row dataset is fast and free, but the combination has real rough edges. An honest assessment after a full Tuva Project run.
-
Deploying a Streamlit analytics app in one afternoon
Streamlit Community Cloud deploys directly from a GitHub repo. Here's the exact setup for a multipage analytics app with a DuckDB backend — including requirements.txt, secrets, and the database file.
-
DuckDB concurrency in 2026: why you can't run dbt and DBeaver at the same time
DuckDB's single-writer model means that opening a .duckdb file in DBeaver blocks dbt from acquiring its write connection. Here's the exact error, why it happens, and the right fix.
-
Feature engineering from claims data for a Random Forest classifier
Healthcare claims have dozens of potential features for patient risk models. Here's how to select and validate features from a Tuva Project dataset for a Random Forest classifier predicting HCC gaps.
-
From raw claims to RAF: what the data pipeline actually looks like
The path from a raw Medicare claim file to a patient's Risk Adjustment Factor score involves five distinct transformation layers. This is what each layer does and where the dbt models fit in.
-
HCC suspecting explained from a data engineering perspective
HCC suspecting is about identifying conditions documented in prior years that haven't appeared in claims yet this year. This is what the data pipeline looks like and what the Tuva mart actually produces.
-
Predicting patient risk with scikit-learn on top of HCC suspecting data
A Random Forest classifier on Tuva's hcc_suspecting__summary table, using age, sex, paid amount, and condition count to predict which patients have above-median HCC gaps.
-
Running the Tuva Project on DuckDB — what breaks and how to fix it
Tuva 0.17.2 on DuckDB 1.10 with dbt 1.11 produces three distinct failure modes. This is what actually broke and how each one was fixed.