Picture by Ideogram
Most of my days as an information scientist appear like this:
- Stakeholder: “Are you able to inform us how a lot we made in promoting income within the final month and what number of that got here from search adverts?”
- Me: “Run an SQL question to extract the info and hand it to them.”
- Stakeholder: “I see. What’s our income forecast for the following 3 years?”
- Me: “Consolidate knowledge from a number of sources, converse to the finance crew, and construct a mannequin that forecasts income.”
Duties just like the above are advert hoc requests from enterprise stakeholders. They take round 3–5 hours to finish and are normally unrelated to the core mission I am engaged on.
When data-related questions like these are available, they usually require me to push the deadlines of present initiatives or work additional hours to get the job achieved. And that is the place AI is available in.
As soon as AI fashions like ChatGPT and Claude have been made obtainable, the crew’s effectivity improved, as did my means to answer advert hoc stakeholder requests. AI dramatically decreased the time I spent writing code, producing SQL queries, and even collaborating with totally different groups for required data. Moreover, after AI code assistants like Cursor have been built-in with our codebases, effectivity features improved even additional. Duties just like the one I simply defined above may now be accomplished twice as quick as earlier than.
Lately, when MCP servers began gaining recognition, I assumed to myself:
Can I construct an MCP that automates these knowledge science workflows additional?
I spent two days constructing this MCP server, and on this article, I’ll break down:
- The outcomes and the way a lot time I’ve saved with my knowledge science MCP
- Assets and reference supplies used to create the MCP
- The fundamental setup, APIs, and providers I built-in into my workflow
# Constructing a Knowledge Science MCP
Should you do not already know what an MCP is, it stands for Mannequin Context Protocol and is a framework that means that you can join a big language mannequin to exterior providers.
This video is a good introduction to MCPs.
// The Core Downside
The issue I wished to resolve with my new knowledge science MCP was:
How do I consolidate data that’s scattered throughout numerous sources and generate outcomes that may instantly be utilized by stakeholders and crew members?
To perform this, I constructed an MCP with three parts, as proven within the flowchart beneath:
Picture by Writer | Mermaid
// Part 1: Question Financial institution Integration
As a information base for my MCP, I used my crew’s question financial institution (which contained questions, a pattern question to reply the query, and a few context concerning the tables).
When a stakeholder asks me a query like this:
What proportion of promoting income got here from search adverts?
I not must look by means of a number of tables and column names to generate a question. The MCP as a substitute searches the question financial institution for the same query. It then features context concerning the related tables it ought to question and adapts these queries to my particular query. All I have to do is name the MCP server, paste in my stakeholder’s request, and I get a related question in a couple of minutes.
// Part 2: Google Drive Integration
Product documentation is normally saved in Google Drive—whether or not it is a slide deck, doc, or spreadsheet.
I linked my MCP server to the crew’s Google Drive so it had entry to all our documentation throughout dozens of initiatives. This helps shortly extract knowledge and reply questions like:
Are you able to inform us how a lot we made in promoting income within the final month?
I additionally listed these paperwork to extract particular key phrases and titles, so the MCP merely has to undergo the key phrase checklist based mostly on the question quite than accessing a whole bunch of pages directly.
For instance, if somebody asks a query associated to “cell video adverts,” the MCP will first search by means of the doc index to determine probably the most related recordsdata earlier than trying by means of them.
// Part 3: Native Doc Entry
That is the best element of the MCP, the place I’ve an area folder that the MCP searches by means of. I add or take away recordsdata as wanted, permitting me so as to add my very own context, data, and directions on prime of my crew’s initiatives.
# Abstract: How My Knowledge Science MCP Works
This is an instance of how my MCP presently works to reply advert hoc knowledge requests:
- A query is available in: ”What number of video advert impressions did we serve in Q3, and the way a lot advert demand do now we have relative to provide?”
- The doc retrieval MCP searches our mission folder for “Q3,” “video,” “advert,” “demand,” and “provide,” and finds related mission paperwork
- It then retrieves particular particulars concerning the Q3 video advert marketing campaign, its provide, and demand from crew paperwork
- It searches the question financial institution for comparable questions on advert serves
- It makes use of the context obtained from the paperwork and question financial institution to generate an SQL question about Q3’s video marketing campaign
- Lastly, the question is handed to a separate MCP that’s linked to Presto SQL, which is robotically executed
- I then collect the outcomes, evaluate them, and ship them to my stakeholders
# Implementation Particulars
Right here is how I carried out this MCP:
// Step 1: Cursor Set up
I used Cursor as my MCP consumer. You may set up Cursor from this hyperlink. It’s basically an AI code editor that may entry your codebase and use it to generate or modify code.
// Step 2: Google Drive Credentials
Virtually all of the paperwork utilized by this MCP (together with the question financial institution) have been saved in Google Drive.
To provide your MCP entry to Google Drive, Sheets, and Docs, you may have to arrange API entry:
- Go to the Google Cloud Console and create a brand new mission.
- Allow the next APIs: Google Drive, Google Sheets, Google Docs.
- Create credentials (OAuth 2.0 consumer ID) and save them in a file referred to as
credentials.json.
// Step 3: Set Up FastMCP
FastMCP is an open-source Python framework used to construct MCP servers. I adopted this tutorial to construct my first MCP server utilizing FastMCP.
(Word: This tutorial makes use of Claude Desktop because the MCP consumer, however the steps are relevant to Cursor or any AI code editor of your alternative.)
With FastMCP, you possibly can create the MCP server with Google integration (pattern code snippet beneath):
@mcp.device()
def search_team_docs(question: str) -> str:
"""Search crew paperwork in Google Drive"""
drive_service, _ = get_google_services()
# Your search logic right here
return f"Trying to find: {question}"
// Step 4: Configure the MCP
As soon as your MCP is constructed, you possibly can configure it in Cursor. This may be achieved by navigating to Cursor’s Settings window → Options → Mannequin Context Protocol. Right here, you may see a bit the place you possibly can add an MCP server. Once you click on on it, a file referred to as mcp.json will open, the place you possibly can embrace the configuration in your new MCP server.
That is an instance of what your configuration ought to appear like:
{
"mcpServers": {
"team-data-assistant": {
"command": "python",
"args": ["path/to/team_data_server.py"],
"env": {
"GOOGLE_APPLICATION_CREDENTIALS": "path/to/credentials.json"
}
}
}
}
After saving your modifications to the JSON file, you possibly can allow this MCP and begin utilizing it inside Cursor.
# Last Ideas
This MCP server was a easy aspect mission I made a decision to construct to avoid wasting time on my private knowledge science workflows. It is not groundbreaking, however this device solves my quick ache level: spending hours answering advert hoc knowledge requests that take away from the core initiatives I am engaged on. I imagine {that a} device like this merely scratches the floor of what is potential with generative AI and represents a broader shift in how knowledge science work will get achieved.
The standard knowledge science workflow is shifting away from:
- Spending hours discovering knowledge
- Writing code
- Constructing fashions
The main focus is shifting away from hands-on technical work, and knowledge scientists at the moment are anticipated to take a look at the larger image and remedy enterprise issues. In some instances, we’re anticipated to supervise product choices and step in as a product or mission supervisor.
As AI continues to evolve, I imagine that the traces between technical roles will change into blurred. What’s going to stay related is the ability of understanding enterprise context, asking the best questions, decoding outcomes, and speaking insights. If you’re an information scientist (or an aspiring one), there isn’t a query that AI will change the best way you’re employed.
You might have two selections: you possibly can both undertake AI instruments and construct options that form this modification in your crew, or let others construct them for you.
Natassha Selvaraj is a self-taught knowledge scientist with a ardour for writing. Natassha writes on every little thing knowledge science-related, a real grasp of all knowledge matters. You may join together with her on LinkedIn or take a look at her YouTube channel.
