Somebody handed us 100,000 queries once and asked what was actually in there. Here's exactly what came out.
We get asked some version of the same question on almost every call: what does Neuron actually do with our query history once we have it. Not the pitch version. The literal, step by step, what happens between "here's a CSV" and "here's your context layer" version.
So let's walk through an actual run. The numbers are real, from a real engagement, anonymized because the client isn't the point. The pipeline is.
Stage one: processing
We start with the raw export. One recent run came in at roughly 100,000 queries, five analysts, three months of work, one production install. Before anything else happens, the pipeline cleans and structures that raw log: dedupes, strips noise, and gets it into a shape where every query can be examined on its own.
That's the boring stage and also the one nobody skips well. Get this wrong and every stage after it inherits the mess.
Stage two: lineage
Next, the pipeline maps how data actually flows. Table to table, column to column, and critically, it starts surfacing the join patterns nobody wrote down. In that same run, we found 184 distinct join patterns, and more than half of them, 109, were non-obvious. Not documented anywhere. Known, if they were known at all, to one or two people.
This is the stage where you find the join that's been running for two years, producing correct-looking numbers, that only one analyst on the team actually understands.
Stage three: catalog
Now the pipeline builds the actual inventory: tables, columns, dimensions, facts, domains. In that run, it surfaced 414 metrics and 4,317 business rules. Not estimated. Counted, directly from the queries, because every one of those metrics and rules was actually run by someone, at least once, in production.
Nobody on that team tagged any of this by hand. The pipeline read the queries and inferred the structure underneath them.
Stage four: KPI analysis
Here's where it gets uncomfortable, in a useful way. This is the stage that scores the practice itself: 46 KPIs across knowledge concentration, efficiency, preservation, and consistency. Of the roughly 17,600 distinct analytical patterns that run surfaced, the pipeline sorted them into tiers. Gold: proven, repeated, multi-use, the real reusable IP, a bit under a third of the total. Silver: one-shot complete queries, working knowledge, about two-thirds. The rest: small reusable fragments, filters and joins other queries build on.
That split alone tells a team something interviews never would: how much of what they've built is actually infrastructure versus how much is a one-time answer to a question nobody will ask again the same way.
Stage five: OSI export
Everything gets combined into one file: a semantic model in Open Semantic Interchange format, one YAML, fully reviewable, ready to plug into whichever platform the team already runs. Not because OSI is the only format that matters, we also output to Snowflake, Databricks, dbt, but because it's the neutral one, the version that doesn't lock anyone into us.
Then: ask your data
Once the model exists, the actual point of the whole exercise shows up. Plain English in, ready to run SQL out, with provenance attached to every line: which neuron it came from, which business rule applied, which metric definition it used. Not a black box answer. An answer you can trace back to the exact query that taught the system what "right" looks like.
Why this takes weeks, not months
The whole thing, assessment through delivery, runs in four to six weeks, with eight to sixteen hours of actual client time. Not because we're rushing. Because every part of it that traditionally takes months (interviewing analysts, reverse-engineering dashboards, writing definitions by hand) doesn't need to happen. The logic already existed. It was sitting in 100,000 queries nobody had read end to end. We just read them.
We're building the context infrastructure to make your analytics memory durable, traceable, and reusable. Implementing AI in your analytics team? Let's talk.