MK
All services
05Data platform & AI engineering

A medallion data platform your finance team signs off, in twelve weeks.

This is for mid-size and enterprise companies whose numbers come from several systems and do not agree. Some have no data team, and finance holds it together with spreadsheets. Some have a data team and a warehouse or lakehouse that nobody trusts, because the cleaning logic is spread across notebooks, dashboards and one person's memory. The medallion pattern fixes that by giving every step a fixed place. Bronze holds each source exactly as it arrived, so nothing is lost and any number can be replayed. Silver is where data is typed, deduplicated and reconciled, and where every business rule is written down. Gold holds the tables your reports and AI tools read, named in your own vocabulary. The run covers one part of the business: one fund, one plant, one product family. That is what makes twelve weeks real rather than a roadmap. The second part of the business is a second run on a platform that already exists.

Abstract cloud data architecture with connected pipelines and storage layers
Start here

Phase Zero: the data edition

Three weeks · Fixed fee, quoted after the discovery call
What you walk away with
  • If you already have a warehouse or lakehouse, a read of how it is layered today and where the business rules actually live
  • Which platform to use: Microsoft Fabric, Snowflake, Databricks, or the one you already have. Whichever suits your company, decided on your data, with the reasons written down
  • One part of the business loaded read-only (one fund, one plant, one product family) and checked line by line against the spreadsheet your team trusts
  • A set of questions written by your own people, run against that data, and scored on how many came back right
  • A list of your systems, spreadsheets and documents: what is clean, what is not, and who owns each one
  • The access model: who can see what, and how that carries through to the assistant
  • A phased plan with a cost, a test for done and a named lead for each phase, plus what your team owns at the end
  • Your options, each with a cost and a time against it. Stopping and staying with your vendor are both on the list
  • If a vendor has already quoted you, I read the quote against the plan and give you plain-English notes

If the paper does not give you a decision you can act on, you do not pay for it.

Phase Zero is the first three weeks of the run. If the paper says do not go ahead, or stay with your vendor, the work stops there. If it says build, the next nine weeks start against the plan.

Need a review before ongoing support?

Phase Zero, the contract terms procurement will ask about, and the other kinds of work I take on.

How I work with larger companies
What this looks like

The twelve-week run

One run, one part of the business, with an outcome at the end of it. Each step has to pass its test before the next one starts.

  1. Weeks 1 to 3
    Phase Zero

    The platform decision and the plan, on paper

    Gate: You decide: build, stop, or stay put

  2. Weeks 4 to 9
    Bronze, silver, gold

    Your sources landed, reconciled and modelled

    Gate: Gold reproduces your existing reports

  3. Weeks 10 to 11
    Assistant

    Your people ask questions, on your chosen model

    Gate: Answers the questions your team wrote

  4. Week 12
    Handover

    Runbooks, data dictionary, your team trained

    Gate: Signed off, then 30 days of warranty

Bronze
  • A lakehouse on Microsoft Fabric, Databricks or Snowflake, on your own tenant, set up as code
  • Each source landed exactly as it arrived, with its origin and load time on every row
  • Incremental loads, so a rerun never duplicates or loses a row
Silver
  • Typed, deduplicated and reconciled tables for each part of the business
  • Every business rule written down, reviewed with the owner of the number, and tested
  • Quality checks that set bad rows aside for review instead of breaking the run
Gold
  • Tables and measures in your own vocabulary: fund, plant, product, customer, period
  • It reproduces the reports you run today before anything new is built on it
  • Access rules set once, used by the dashboards and the assistant alike
AI access
  • Your choice of model, connected to the shared model for questions about numbers: Claude, OpenAI, or an open-source model such as Qwen or DeepSeek, run on your own infrastructure
  • Your data stays in your tenant. Nothing is sent to a model you have not approved
  • Search over contracts, leases, reports and memos for questions about documents, with the passage quoted
  • A test set built from your own questions, run on every change, so accuracy is measured rather than claimed
Handover
  • Runbooks, design notes and the data dictionary, written for the team that takes it on
  • Your engineers paired with me on the connectors and the deployment before I leave
  • A named owner on your side, and a handover signed off in writing
Architecture and data flow diagrams on a workstation screen

Medallion architecture is simple to draw and easy to get wrong. The usual failure is a gold layer built straight from raw extracts, with the business rules hidden inside dashboard formulas. The reports look right until two of them disagree, and then nobody can say which one is wrong. I build it the other way round. Every rule that changes a number lives in silver, written down and reviewed with the people who own that number. Gold only shapes and aggregates. A figure in a board report can be traced back through gold and silver to the bronze row and the source file it came from.

The order matters, and so does the test. Silver and gold are not done when the pipelines run. They are done when they reproduce an agreed set of the reports you already trust, within a tolerance agreed before the work starts. Most of the early weeks go on finding out why three systems give three answers to the same question. Each difference becomes a rule. When the reports match, those rules are your definitions, not one person's memory. Only then does anything new get built on top, whether that is a dashboard, a forecast or an assistant. An assistant over unchecked data gives confident wrong answers, and the first one a director catches is the last time anyone uses it.

Your engineers own the source connectors and run the platform after handover. While I am there, I own the layer design, the silver rules, the gold model and, if you want one, the AI layer. Everything is set up as code with a deployment pipeline, so a change to a rule is reviewed like a change to software. Access is set once in the platform, and the dashboards and the assistant both follow it. The handover is written so another supplier could pick it up without calling me. I am building this now for a US real estate investment manager: a medallion lakehouse on Microsoft Fabric with an assistant on top. That work is described on this site as far as the client allows.

Relevant engineering experience

These projects show the engineering work behind the service. Each case study describes its own scope and outcomes.

Bring the question the board asked that nobody could answer quickly, and the systems the answer lives in. The first call establishes whether Phase Zero is the right next step.

How it runs
  1. Weeks 1–3

    Phase Zero

    • Interviews with the sponsor, the technical owner and two of the people who build the reports today
    • One part of the business loaded read-only and checked against the spreadsheet you trust
    • Your own question set run and scored
    • The paper: platform decision, options with a cost and a time, and a phased plan
  2. Weeks 4–9

    Bronze, silver and gold

    • The lakehouse on your tenant, set up as code from the first week
    • Sources landed in bronze in priority order, each one checked before the next starts
    • Silver rules agreed with finance in the room, written down and tested
    • Passes when gold reproduces the agreed reports
  3. Weeks 10–11

    AI access

    • The assistant connected to the shared model, read-only, with the same access rules
    • Document search over the agreed stores, with the passage quoted on every answer
    • A test set from your questions, run on every change
    • Passes on the score against your own questions
  4. Week 12, plus 30 days

    Handover

    • Your engineers paired on connectors and deployment before I leave
    • Data dictionary, runbooks and ADRs for the team that inherits it
    • Thirty days of warranty on what was accepted, window and cap in the SoW
    • The second domain, if there is one, is a second run on a platform that now exists
What you walk away with
  • A medallion lakehouse (bronze, silver, gold) on your own tenant, set up as code with a deployment pipeline
  • Silver rules and gold measures that reproduce your existing reports, signed off by finance
  • An assistant over your numbers and documents, with sources shown and a test set
  • The data dictionary, runbooks and design notes your team owns after I leave
Tools & tech
Microsoft Fabric
Snowflake
Databricks
PySpark / Spark SQL
Delta Lake / Apache Iceberg
dbt
Power BI
Claude / Anthropic API
OpenAI API
Open-source models (Qwen, DeepSeek, Llama) via vLLM / Ollama
Azure / AWS
Terraform
PostgreSQL / pgvector
How we engage

One twelve-week run for one part of the business. Three weeks to the Phase Zero paper. Six weeks to a platform that passes its test. Two weeks to an assistant that passes its test. One week of handover, then thirty days of warranty. Each phase is a fixed fee with a written test for done, quoted after Phase Zero. You can stop at the end of any phase. A second part of the business is a second run.

What I guarantee

Done means it works on your own numbers. A demo does not count.

The run is twelve weeks and the outcome is inside it. By week twelve the bronze, silver and gold layers for one part of your business are live on your own tenant. Each phase has a written test for done, agreed before it starts. You pay when a phase passes that test, not on a date. The platform passes when gold reproduces an agreed set of your existing reports. If there is an assistant, it passes when it answers the questions your team wrote, and every answer shows where it came from.

  • Fixed fee per phase. You pay when the phase passes its test.
  • Read-only throughout. Nothing writes into a core system without its own test and sign-off.
  • You can stop at the end of any phase and pay only for the phases that passed.
  • A defect warranty on each phase. The length and the cap are set in the statement of work.
  • You own the platform, the pipelines and the data model. I keep the tools I built before we started.
  • A short standard contract plus a one-page statement of work, or your contract if you have one. UK and US paper, invoiced in USD.
Questions

Questions I get asked about this.

What is medallion architecture?

A way of layering a data platform so every step has a fixed place. Bronze holds each source exactly as it arrived. Silver holds the same data typed, deduplicated and reconciled, with every business rule written down. Gold holds the tables and measures your reports and AI tools read. The value is that a wrong number can be traced back layer by layer to the rule or the source row that caused it, instead of being argued about in a meeting.

We already have a lakehouse and nobody trusts it. Can you fix that?

Yes, and it is a common starting point. Phase Zero reads how the current platform is layered and where the business rules actually live. Often they sit in dashboard formulas and notebooks rather than in one place. The paper says whether to restructure what you have into bronze, silver and gold, or to rebuild one part of the business alongside it and move across. Keeping your current platform is always on the list of options.

Microsoft Fabric, Snowflake or Databricks?

We decide in Phase Zero, on your data. If you already use Microsoft and Power BI, Fabric is usually the answer. If you run more than one cloud, share a lot of data with outside firms, or already have Snowflake, Snowflake is usually the answer. If you have heavy engineering or machine learning work, Databricks may suit you better. Sometimes the right call is to keep what you have and fix the model on top. If so, the paper says that.

Which AI model do we have to use?

Your choice. I build the data layer so any model can sit on it. That can be Claude, OpenAI, or an open-source model such as Qwen or DeepSeek deployed on your own infrastructure, so your data never leaves your tenant. If your company already runs Copilot or Azure OpenAI, the same layer serves that too. Claude is my default because it handles long documents and tools well, but the choice is made in Phase Zero and written into the access model.

Is this a RAG project?

Half of it. Questions about documents are answered by searching the documents and quoting the passage. That is RAG. Questions about numbers are answered by querying the shared model of the business. That is a different method, and it is the one most boards want. A project that stops at RAG leaves the numbers unanswered.

Do we need a data team?

No. Some clients have one and some do not. Your engineers own the source connectors and run the platform after handover. While I am there, I own the layer design, the silver rules, the gold model and the AI layer. If you have a data team, they pair with me throughout and own all of it at the end. If the paper finds you need to hire someone to run it, it names the role.

How is security handled?

Access rules are set once in the platform. The dashboards and the assistant both follow them. The assistant is read-only and cannot write into any of your systems. Everything runs on your own tenant and your own accounts.

How long does it take?

Twelve weeks for one part of the business. Three weeks of Phase Zero. Six weeks for bronze, silver and gold. Two weeks of assistant. One week of handover. The date holds because the scope is fixed in Phase Zero: one part of the business and one question set. What moves the date is how fast your side gives access and makes decisions.

Why only one part of the business? We have the whole company to do.

Because the whole company in twelve weeks is a roadmap, not an outcome. One fund, one plant or one planning process is enough to prove the platform and the assistant on real questions. The next part of the business is a second run. It is faster, because the platform already exists.

What does it cost?

The run is a fixed fee, split into phases. Each phase has a written test for done. I quote it after Phase Zero, once I have seen your sources and systems. The Phase Zero fee is set on the first call. You pay when a phase passes its test, not on a date.

One set of numbers, and a platform that shows where each one came from.

A first call establishes the decision in front of the steering group, who signs, and which systems the answer lives in. That is what sets the Phase Zero fee. No rate card, and nothing to put out to tender at this size.