Every team standing up AI capability inherits a decades-old modeling debate and asks which pattern wins — Inmon's normalized core or Kimball's dimensional star. Framed as a contest, it is a false binary: the two were built for different jobs, and AI workloads span both.
This quarter's briefing weighs the evidence inside the DMBOK frame and lands on a defensible answer — a normalized, conformed atomic foundation feeding a dimensional serving layer. Not a compromise between the two camps, but the order modern lakehouse practice already follows.
Framed as a contest, “3NF versus dimensional for AI” is a false binary. Weighed against the DMBOK frame, the honest answer is a layered one — a normalized core feeding a dimensional serving layer.
Framed as a contest, the question is a false binary — the evidence points to a normalized core feeding a dimensional serving layer.
Read · 5 minGrain, flexibility, and integrity — why an atomic, integrated core is the sounder base for ML and LLM workloads.
Read · 4 minTreat every metric like a load-bearing function — write the expected number first, then assert it against mocked rows.
Read · 5 minMeaning, structure, storage — why the logical layer is the indispensable bridge between them.
Read · 4 minObservability, idempotency, and the small operational disciplines that separate a POC from a platform.
Coming soonWhen streaming, when hub-and-spoke, and when the honest answer is a scheduled batch job.
Coming soonOne named engagement, plus representative stories attributed to industry only, at our clients’ request.
ONE Solutions, a roofing-industry company, needed a data environment it could actually run itself. IDS re-platformed it from Azure to Snowflake, built the pipelines in dbt, wired up the integrations, and modeled the warehouse with data-quality controls baked in — planning heavily upfront to minimize rework and keep the build simple, maintainable, and cost-effective.
Re-platforming cut cloud spend to a third, dropped the monthly platform bill from $250 to $71.92, and took pipelines that once ran 3–6 hours down to under 30 minutes — roughly $30K of year-one value and $10K+ in recurring annual savings. Through documentation, in-person training, and knowledge transfer, the client’s own data engineer was left fully confident managing the environment IDS built.