Private beta — onboarding data teams now

An end-to-end data platform, run by AI agents

From source system to trusted answer.

AI agents do the data engineering. Your team approves every change.

NerveStax builds and operates your ingestion, models and semantic layer on the stack you already own, and serves answers your business can rely on.

From source system to trusted answerSOURCE SYSTEMSSAPPostgreSQLMySQLA PERSON APPROVES EACH STAGEIngeston your AirflowModelbuilt and testedOperatewatched, repairedDefineone meaning per numberANSWERRevenue last week?$795.3K▼ 7.1%source citedRUNS ON WHAT YOU OWNYour gitYour AirflowYour warehouseYour AI model
  1. Ingest

    Source systems are loaded into your warehouse by jobs on your own Airflow.

    A person approves the load plan and merges the install.
  2. Model

    dbt models are written, built and tested in an isolated sandbox.

    A person reviews and merges the pull request.
  3. Operate

    Schedules, failures, alerts and backfills are watched and worked.

    A person approves what the agent runs in production.
  4. Define

    The semantic layer: each measure defined once, with its access rules and tests.

    A person merges every definition.
  5. Answer

    Conversational analytics, dashboards, data apps, your BI tool and AI assistants.

    Every query runs as the person asking.
Runs on what you own
  • Your git server
  • Your Airflow
  • Your warehouse
  • Your AI model
  • In productionsince August 2026
  • On your stackyour git, Airflow and warehouse stay where they are
  • Every changearrives as a pull request a person merges
  • Sovereign AIdeploy in your environment, on a model you host

The business is asking for answers, and for AI. Your team is busy keeping the pipelines running.

The backlog grows

Every new source, model and fix waits for the same few engineers.

Numbers stop agreeing

Each team, and each AI assistant, works out revenue its own way, and the answers differ.

AI waits on the data

The AI plan depends on data that is governed, documented and current.

The product

One application, from ingestion to answers.

The same workspace serves the engineer, the analyst and the business user. These are its screens.

Conversational analytics: A question in plain language, answered with cited figures and a chart.
Conversational analyticsA question in plain language, answered with cited figures and a chart.
Dashboards: Built over the same definitions. Every query runs as the viewer.
DashboardsBuilt over the same definitions. Every query runs as the viewer.
Data apps: A weekly story for leadership, written from governed numbers.
Data appsA weekly story for leadership, written from governed numbers.
Explorer: A slice-and-dice data app for analysts, on the same numbers.
ExplorerA slice-and-dice data app for analysts, on the same numbers.
Ingestion: Source systems loaded into your warehouse, each with its state and last run.
IngestionSource systems loaded into your warehouse, each with its state and last run.
The agent: An engineer asks for a change. The agent plans it, asks for approval and reports back.
The agentAn engineer asks for a rebuild. The agent plans it, waits for approval, runs it and reports the result.
Audit log: Who did what, to which object, when and with what result.
Audit logWho did what, to which object, when and with what result.
AI models: The model accounts your organisation connects, commercial or self-hosted.
AI modelsThe model accounts your organisation connects, commercial or self-hosted.

01 · Answers the business can trust

One definition of every number, wherever it is asked.

The semantic layer holds each measure once: what it means, how it is calculated and who may see it. Every way of asking reads from that definition, so the answers agree.

Semantic layer

Defined once. Served five ways.

Conversational analytics, dashboards, data apps, the BI tool you already run and your AI assistants all read the same certified definition. Nothing is rebuilt and nothing is recalculated on the side.

  • Checked before it is served. A definition that fails its checks is not published.
  • Tested on a schedule. Questions with expected results run after every change and every night.
  • Adds to your BI. A PostgreSQL-compatible connection, and connectors for Claude, ChatGPT and Gemini.

The semantic layer in detail, with a film →

One definition, the same number everywhereDEFINED ONCERevenueTESTED · GOVERNED$795.3KCONVERSATION$795.3KDASHBOARD$795.3KDATA APP$795.3KYOUR BI TOOL$795.3KAI ASSISTANT

Conversational analytics

The AI model never writes a number.

Every figure in an answer comes from a query against a certified definition and is shown with its source. A sentence whose number did not come from a query is removed. When no definition covers the question, the reply says so.

Every figure comes from a queryTHE ANSWERRevenue was$795.3K, down7.1%on last week.Each figure is tied tothe query it came from.QueryCERTIFIED DEFINITIONIt should reach $900K next week.No query behind it.Removed before you see it.

Governed access

One question, the right answer for each person.

Row and column rules are written once, beside the definition, in plain sentences. They apply on every route: a question, a dashboard, a BI tool or an AI assistant.

  • Rows filtered by group or by the viewer’s own attributes.
  • Measures and dimensions hidden from chosen roles.
  • Each person can see which rules apply to them.
One question, three people, three views“Revenue by country?”Finance leadAll countriesMargin shownUK managerUK onlyMargin hiddenYour customerOwn rows onlyEmbedded view

02 · AI does the engineering; your people approve

Sources, models and operations, delivered for your review.

Engineers describe what they need in a conversation. Agents do the work in an isolated sandbox and hand back something a person can check: a plan, a pull request or a verdict. How the agents work →

Ingestion

New sources, loaded on your own Airflow.

The agent explores the source, and rules in the platform choose how each table loads from the evidence: merge, append or full refresh. You review the plan, and it installs as a pull request.

  • Sources today: MySQL, PostgreSQL and SAP, including S/4HANA, Business One, ByDesign and SuccessFactors.
  • Passwords and keys go into a secure form and never reach the AI model.
  • A changed column type stops that table before anything is written.

Ingestion in detail, with a film →

Rules choose how each table loadsSAPORDERSEVIDENCEUnique keyUpdate timeNull ratesRULES CHOOSEMergeAppendFull refreshInstalls as a pull requestYOU REVIEW · YOU MERGE

Modelling

Model changes arrive as tested pull requests.

The agent edits your dbt project on its own branch, builds and tests only what the change affects, and opens a pull request under the engineer’s own identity. Each pull request is then tested on your Airflow in an isolated schema.

  • Merge checks stop a change that would break a schedule or a certified definition.
  • The pull request carries its build and test results.
  • Merging always happens in your git server, by a person.

Tested before merge, in detail, with a film →

A tested pull request, merged by a personMAINPRODUCTIONTHE AGENT'S BRANCHEditBuildTestPull requestTests passed · 4 models touchedYou merge

Operations

Failures are triaged before anyone is paged.

Each alert from Airflow or dbt opens a triage by the on-call agent, which ends in one of two verdicts: resolved, with an explanation, or escalated, with a cause. Backfills are planned by the platform, run in chunks and verified afterwards.

  • Triage is read-only: the agent cannot change anything while it investigates.
  • A backfill is checked to confirm that rows outside its window did not change.
  • A workspace kill switch stops backfills at once.

On-call in detail, with a film →

Backfilling in detail, with a film →

Every alert ends in one of two verdictsAlert02:14TriageREAD-ONLYResolvedwith an explanationEscalatedto you, with the cause

03 · Governed and recorded

Nothing reaches production without a person.

Approval is part of the platform, not a setting someone can forget. The agent has only the permissions of the person it works for, and what it does is on the record.

These always wait for a person

  • Opening a pull request
  • Merging it, which happens in your git server
  • Installing, running, stopping or re-syncing an ingestion
  • Triggering, pausing or clearing an Airflow job
  • Production backfills beyond the limits you set
  • Publishing a dashboard
  • Reading a stored secret

The approver is recorded. Approving an action needs the same permission as doing it.

Work stops at a person before productionTHE AGENT PREPARESPull requestBackfillIngestion installYou decideProductionRecordedWHO · WHAT · WHEN
  • The agent acts as the personIt has only that person’s permissions, re-checked on every turn.
  • Isolated by designWork runs in a sandbox with no git token inside.
  • On the recordMore than 90 kinds of action in an audit log that cannot be edited or switched off.
  • You set the limitsPolicies decide what is small enough to start without sign-off.

04 · Sovereign AI

Sovereign AI, on the stack you own.

Deploy NerveStax entirely inside your own environment and run it on an AI model you host. Your code, your data and your prompts stay where you decide.

Sovereign AI: NerveStax inside your environmentYOUR ENVIRONMENTSOVEREIGN AIYour gitCODEYour AirflowJOBSYour warehouseDATAYour AI modelHOSTED BY YOUNerveStaxINSTALLED HERECode, data and prompts stay inside your boundary.
  • Deploy it where you decide

    Use the service we host, or install NerveStax in your own Kubernetes environment, in your cloud or your data centre.

  • Bring your own AI model

    An open-weights model on your own hardware, through any OpenAI-compatible endpoint, or a commercial model with your own key: Anthropic, OpenAI or Azure OpenAI.

  • Your code and data stay yours

    Code lives in your git and changes only through pull requests. Jobs run on your Airflow. Data stays in your warehouse.

  • No dependence on the largest models

    The platform makes its decisions in code and leaves the wording to the model, with separate models for engineering work and for business questions.

  • Single sign-onAny identity provider that supports OIDC. Sign-on can be enforced.
  • RolesOwner, admin, engineer, analyst and consumer, checked on every screen and every agent tool.
  • Audit logExport and retention you set.
  • SecretsEncrypted at rest and never shown to the AI model.

Security in detail →

Supported today

What it works with.

Ingestion
From MySQL, PostgreSQL and SAP over OData. Lands in MotherDuck today.
Warehouses
Trino, ClickHouse, DuckDB and MotherDuck, validated end to end. Others are in validation.
Transformation and orchestration
dbt, on Apache Airflow 2 and 3.
Git
GitHub and GitLab, cloud and self-hosted.
AI models
Anthropic, OpenAI, Azure OpenAI, or any OpenAI-compatible endpoint.
Identity
Any OIDC provider.

For consulting and delivery partners

Deliver more with the senior people you have.

Talk to us about partnering
  • Agents build, your consultants reviewThe build work is done by agents. Your people set the direction and approve the result.
  • Your quality bar is enforcedEvery change is a tested pull request under the consultant’s own identity, with a record of who approved it.
  • The client owns the resultEverything lands in the client’s git, Airflow and warehouse, so handover is clean.
  • August 2026In production since
  • Every weekNew releases. See what shipped
  • 90+Kinds of action in the audit log
  • DubaiWhere NerveStax is built

See it on your own repository.

Tell us what you run and what is waiting. We will set up a working session on your stack, not a slide deck, and reply within a working day.

→ [email protected]

What’s this about?