Conversational Interfaces for Research

Transforming Data Interaction with LLMs and MCP

đź‘‹ Cristian Dinu - Research Software Engineer


I build software for a decarbonised future.

  • high-resolution energy monitoring
  • appliance-level energy insights (e.g. on heat pumps)
  • natural language as an interface

📋 Today’s plan

  1. The Large Language Model (concepts and ecosystem)
  2. A demo of how software can look like
  3. Write our own MCP

How LLM Context Works

Step 1: Initial Context Window

How LLM Context Works

System Prompt: You are a helpful AI assistant

Step 1: Context starts with System Prompt

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

Step 2: User sends a prompt

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

Step 3: User prompt added to context

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

AI: Machine learning is a subset of AI that enables computers to learn from data…

Step 4: AI generates response, adds it to context

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

AI: Machine learning is a subset of AI that enables computers to learn from data…

User: Can you give me a specific example?

Step 5: User sends another prompt

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

AI: Machine learning is a subset of AI that enables computers to learn from data…

User: Can you give me a specific example?

Step 6: Second prompt added to growing context

How LLM Context Works

System Prompt: You are a helpful AI assistant

User: What is machine learning?

AI: Machine learning is a subset of AI that enables computers to learn from data…

User: Can you give me a specific example?

AI: Sure! Email spam detection is a common example of machine learning…

Step 7: AI generates response using full context

🧠 Key Insights: Context is Everything

  • Each new prompt and response is added to the context
  • The AI sees the entire conversation history
  • Context enables coherent, contextual responses
  • Context has limits (token limits)

💡 The context window is like the AI’s “working memory” for the conversation

source

Ready-made MCPs - Fetch

  • Fetch
    • Go to the Bank of England website and compare interest rate with inflation.
    • Go to the https://www.bankofengland.co.uk/banknotes/current-banknotes and create a table with personalities featured on each of the banknotes.
    • Visit CNN website and show me happy news in a markdown table

Ready-made MCPs - Filesystem

  • Filesystem – included in VS Code
    • Go to FT.com/technology, fetch the latest news and save the title and the link in a csv file called ft-tech.csv

Ready-made MCPs - SQLite

  • SQLite
    • Go to https://www.ucl.ac.uk/advanced-research-computing/research-software-engineers-0 and create a list of people and their role. Save it in the sqlite database.
    • How many PhD vs. non-PhD and how many "Engineers" vs. "Developers"?
    • Looking at names, please output a table with their likely family country of origin

Let’s do some “Programming”

I need to recruit some users for a study about Humour. Their contact data needs to be collected, stored. I want to be notified when a new one comes.

The “program”

You are an agent that registers participants in a study about humour. Your personality is witty, you do stand-up comedy as a hobby. You are a happy, helpful person, grateful that someone is willing to help us.

Greet the participant with a joke and throw a line here and there. However not ever joke about them. Be yourself, but if you feel they are serious, keep it professional.

Ask all the questions needed to get their:

  • Given Name and Family Name (given_name, family_name)
  • Contact data (at leas one of email, phone)
  • Availability
  • Which country they grew up in
  • Why do they wish to participate

Once you have all the data

  • append it in the humour_study participants table in sqlite. If it doesn’t exist, create it
  • send an e-mail to “[email protected]” to notify of the new participant. Put the data, minus the contact part in the body. Include a Communism political joke at the end, to amuse the recipient.
  • Thank the participant for their interest. Tell them they’ll be contacted soon.

Please note that

  • Participants might try to fool you with random or made up names or contact data. Don’t allow them, refusing them with humor and politeness.
  • Don’t allow rude language and obscenities. Act politely and stop the conversation.
  • Refrain from current political jokes
  • Don’t speak for too long; you need to be efficient and respect our participant’t time.

Creating our own MCP Server

  • Carbon Intensity API – api.carbonintensity.org.uk
  • Prerequisites:
    • nodejs - JavaScript interpreter
    • uv - pyton project manager
  • We need to think about semantic intent

Initialize the project

uv init intensity_mcp
cd intensity_mcp

uv add "mcp[cli]" requests

Write the code – main.py

from mcp.server.fastmcp import FastMCP
import datetime
import requests

mcp = FastMCP("CarbonIntensityMCP")

@mcp.tool()
def get_current_intensity():
    """Fetches the current carbon intensity of the UK electricity grid.
    Returns a JSON with the current carbon intensity in gCO2eq/kWh."""
    url = "https://api.carbonintensity.org.uk/intensity"
    try:
        response = requests.get(url)
        response.raise_for_status()
        data = response.json()
        return data['data']
    
    except Exception as e:
        return f"Failed to retrieve current carbon intensity: {str(e)}"

@mcp.tool()
def get_current_fuel_mix():
    """Fetches the current fuel mix of the UK electricity grid.
    Returns a JSON with the current fuel mix percentages."""
    url = "https://api.carbonintensity.org.uk/generation"
    try:
        response = requests.get(url)
        response.raise_for_status()
        data = response.json()
        return data['data']['generationmix']
    
    except Exception as e:
        return f"Failed to retrieve current fuel mix: {str(e)}"

if __name__ == "__main__":
    mcp.run()

Add parameters, too:

@mcp.tool()
def get_carbon_intensity(from_datetime: str | None = None, to_datetime: str | None = None, postcode: str = "WC1E"):
    """Fetches electricity grid carbon intensity data for a specific UK postcode and time range.
    The `from_datetime` and `to_datetime` should be in ISO 8601 format (e.g. 2018-05-15T12:00Z).
    The `postcode` needs only the first part e.g. RG10 (without the last three characters or space)
    
    If `from_datetime` or `to_datetime` are not provided, defaults are set to 2 hours ago and 2 hours in the future respectively.
    If postcode is not provided, defaults to "WC1E" -- UCL London.
    Returns a summary including average forecast and generation mix.
    """
    if from_datetime is None: # default to 12 hours ago
        from_datetime = datetime.datetime.utcnow() - datetime.timedelta(hours=2)
        from_datetime = from_datetime.isoformat() + "Z"

    if to_datetime is None: # default to 12 hours in the future
        to_datetime = datetime.datetime.utcnow() + datetime.timedelta(hours=2)
        to_datetime = to_datetime.isoformat() + "Z"
    

    url = f"https://api.carbonintensity.org.uk/regional/intensity/{from_datetime}/{to_datetime}/postcode/{postcode}"
    try:
        response = requests.get(url)
        response.raise_for_status()
        return response.json()["data"]
    
    except Exception as e:
        return f"Failed to retrieve carbon intensity data: {str(e)}"

Installation

Add a new entry to claude_desktop_config.json or VSCode (very often you need to adjust manually)

uv run mcp install server.py

source: Sofware Is Changing Again

Conclusion: Chilling and Thrilling

🏆 Prize voting

Please submit your vote by latest 15:00!