MGMT +
CNSLTNG

The agentic data stack · Sydney, Australia

Ask your own data anything.
Own the answer.

Agents on your warehouse, your models, your servers. Every answer arrives with the query behind it, and every conversation stays yours.

ClickHouse · MCP · LibreChat · Langfuse, self-hosted

Start with a Read See the layers

On your infrastructure

The conversation LibreChat
The data ClickHouse with MCP
The record Langfuse
Your servers. Your keys. Your traces.

The questions arrive faster than the answers.

The numbers sit wherever the business has put them over the years: point of sale, accounting, payroll, commerce, spreadsheets, inboxes. Nobody trusts all of them, so someone builds a spreadsheet, then a dashboard, then a chart nobody opens, and by the time the answer exists the decision has been made. The architecture that changes this is open source, and large firms already run it. What is missing at your size is someone who can put it in, make your data arrive correctly, and keep it running.

Three layers, one conversation

All of it open source, all of it on your own infrastructure. Each layer does one job.

01 · THE CONVERSATION

A chat window on your own network. Your choice of model, from a frontier model with your own key to one running on your own hardware, and the conversation stays in your own database.

LibreChat

02 · THE DATA

A warehouse built to answer questions, with an agent interface in front of it. The agent writes the query, the warehouse runs it, and the answer comes back as a table or a chart. Nobody has to write SQL.

ClickHouse with MCP

03 · THE RECORD

Every prompt, tool call, cost and latency traced and scored. This is what makes an answer auditable rather than anecdotal, and it is how the answers improve.

Langfuse

The part nobody sells you

Open source gives you the layers. It does not give you your data.

The work is everything before the warehouse: finding the source systems, matching the identifiers that disagree, landing the history, reconciling the totals against the source, and knowing the moment a feed stops. That part is not in a repository, and it is what makes the answers believable.

Your systems

  • Point of sale
  • Accounting
  • Commerce
  • Payroll
  • Spreadsheets
  • Email and documents

The pipe

  1. Extract
  2. Match
  3. Land
  4. Reconcile

The warehouse

One place every number comes from, and one place to ask questions of.

The answers

Tables, charts and briefs, each traced back to the rows behind it.

Every number traceable to a source row, and a daily report of what arrived and what did not.

Proof in action

One demonstration on synthetic data: the question as a person would ask it, and the machinery behind the answer.

Which sites ran labour above 30 per cent of sales this month?

Illustrative example. Synthetic data.

The question, and the answer

Which sites ran labour above 30 per cent of sales this month?

Three sites are above it. Ashfield is the highest, at 42 per cent of sales.

Ashfield 42%
Bondi 36%
Newcastle 33%

The answer arrives as a chart and a table. No dashboard to build, and no analyst in the queue.

The query, and where the numbers came from

SELECT site,
       sum(sales) AS sales,
       sum(labour_cost) AS labour_cost,
       round(sum(labour_cost) / sum(sales), 3) AS labour_ratio
FROM pos.sales_daily
WHERE month = '2026-03'
GROUP BY site
HAVING labour_ratio > 0.30
ORDER BY labour_ratio DESC

Written by the agent, run against your warehouse. The prompt, the query, the rows and the cost of the answer are all recorded, so any figure in this page can be checked later.

Yours to own

  • Your data Stays on your infrastructure, in your own account.
  • Your conversations Stay in your own database, not in someone else's.
  • Your traces Belong to you, not to a SaaS product.
  • Your model choice Including one that never leaves the building.

No per-seat subscription. No per-query billing. It does cost a server and someone to look after it, and that is the retainer.

Who it is for

For

  • A firm with real data in several systems and no data team.
  • A technical team that wants it built, then taken over.
  • An enterprise that wants a private pilot before a platform decision.

Not for

  • Anyone shopping for the cheapest dashboard.
  • A public chatbot with your logo on it.
  • Anyone who cannot let one person own their data.

How it starts

The same three depths as the rest of my work, in the order this work happens.

  1. 01 · THE READ

    A data audit. What sources exist, who owns each one, where the numbers disagree and by how much. Days, not months, with every number sourced.

  2. 02 · THE BUILD

    The pipe and the stack. Data arriving on its own, agents on top of it, all of it on your infrastructure.

  3. 03 · THE STAY

    Running it. New sources, broken feeds, reconciliation reports, and a response time we agreed in advance.

Agreed before any work starts

Which source systems are in scope. How often the data arrives. How quickly I respond when something breaks. Everything else is deliberately out of scope, in writing.

Start with a Read.

Tell me what you have and what you cannot see. Days later you will know what your data can answer, and what it would take.

Thank you.

The problem is with me now. I'll be in touch.