Strata launches a semantic layer that refuses invalid LLM queries, built by a four-year Netflix self-service analytics engineer
Strata, a new semantic layer and dashboards product in free Docker beta from an ex-Netflix engineer, pitches itself as the layer an LLM has to talk through, with the right to refuse a plausible-looking but invalid query.
Image: Strata Business Intelligence
Why it mattersA team shipping a natural-language analytics assistant can keep the agent on a leash by writing the measures and grains in a semantic layer that refuses the queries that would double-count or cross grains.
An LLM analytics assistant that returns the wrong number once is a story the business will tell for a year. Strata, posted to Hacker News on 30 September 2026, is a semantic layer and dashboards product whose pitch is that the layer gets the right to refuse an invalid query before the model runs it.
The author, posting as ajoski9, writes in the Show HN thread that he spent four years at Netflix on self-service analytics for non-technical users. He describes Strata as "an expressive semantic layer that can say no to your LLM", and in a follow-up comment says a semantic layer "rejects plausible queries that are invalid" and keeps agents from writing problematic SQL. The Hacker News submission carries 22 points and 15 comments at the time of writing.
What it refuses, and how
The product homepage describes grain-safe blending that aggregates every fact table to a common grain before joining on conformed keys, "never fact to fact". The practical consequence is that a measure crossing two fact domains is composed through conformed dimensions rather than a direct join, which is how double-counts enter a dashboard. Partition and aggregate-aware semantic routing picks which compute tier answers a query, across ClickHouse, Druid, Snowflake and AWS Athena in the OLAP engines the site lists.
Strata ships five measure types: Standard, Complex, Snapshot and two Level of Detail variants, plus segments for cohort analysis, row-level security built into the model, report subscriptions and a Google Sheets add-on. Those are the product's own feature list, not an independent capability review.
What is available today
The homepage offers a free beta on Docker for up to 25 users, with a 15-minute setup. Strata is deliberately not open source. The company is Strata Business Intelligence, LLC, per the footer. One commenter on the Show HN thread flagged documentation 404s, which the author acknowledged.
The customer story on the homepage is Ajo's Netflix work: a customer service organisation with 800 non-technical users, weekly self-serve analytics users went from 5 to over 100, and ad-hoc data requests dropped by 90 percent. Those are the company's own numbers for a prior employer, not a measurement Strata has published on itself.
The idea the item carries is independent of whether Strata is the right vendor to buy. A team building a natural-language analytics feature of its own has two options for stopping the agent from inventing numbers. One is to write the measures, grains and allowed joins down as a semantic layer that the model has to query through, so the layer rejects anything it cannot answer. The other is to let the model write SQL and hope the eval suite catches a cross-grain double-count on the first run. The first is the harder project upfront and the one the Strata team has picked a public fight for.
Source
- Primary source: Strata homepage
- Related: Show HN discussion, 22 points, 30 September 2026
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
