Documentation
NemulAI tells you which customers are unprofitable while their price can still change — by attributing real inference cost to the account that caused it and comparing it to what you billed. No credentials, no connector, no row of your data.
Quickstart
- 1Open the audit flow — no account needed.
- 2Copy the generated metadata query, run it in your warehouse, paste the result back.
- 3Check the mapping and the readiness score. Both steps so far are free.
- 4Run the generated read-only export SQL and upload the results.
- 5Read margin per customer, worst first — example.
How an audit works
- 1
Verify your schema
Pick your warehouse. We generate a metadata-only query — table names, column names, data types, and declared keys where the catalog exposes them. You run it yourself and paste the result back. NemulAI never connects to your database.
Free. Anonymous. Nothing is stored unless you choose to save it.
- 2
Review the mapping and readiness
Every column is mapped onto our canonical economics model with a confidence label, and every guess is correctable with a dropdown. The readiness score states what the audit can and cannot compute from your schema as mapped — including what is missing and why it matters.
Free. You see the gaps before you pay, not after.
- 3
Run the audit
We generate read-only, column-limited, time-bounded extraction SQL for your dialect. You run it, export CSVs, and upload them. We attribute cost per request, join revenue, and produce per-customer margin worst-first — plus a PDF.
Paid. A one-off 7-day audit that converts to monthly monitoring.
The short version of the security model
We receive
- ✓Table and column names
- ✓Data types and nullability
- ✓Declared keys, where your catalog exposes them
- ✓Aggregated usage and billing rows you export and upload
We never receive
- ✕Database credentials or connection strings
- ✕OAuth grants, hostnames, or network access
- ✕Prompts, completions, or message content
- ✕Any row we did not ask you to export
Margin audit guide
Security & data handling
What leaves your environment at each step, what we never receive, and how each boundary is enforced rather than promised.
Schema metadata format
The metadata query we generate per warehouse, the eight columns it returns, and the paste and CSV formats we accept back.
Generated export requirements
The four exports an audit runs on — usage, revenue, customers, infrastructure — which are required, and what each one unlocks.
Questions
No tenant id, no per-request cost, what happens to unattributable cost, and what becomes of uploaded exports.
Optional continuous monitoring — advanced
The audit needs none of this. These cover the optional agent for customers who own their GPUs and want measured device time rather than modelled cost, after an audit has already told them where to look.
Legal documents: Privacy policy · Terms · DPA · Security questionnaire