Data analysis agents & project scoping

Part 1 · Theory (30 min)

Deterministic vs. agentic systems

You now know how to build an agent — but the most senior move in engineering is knowing when not to. Some jobs want a plain, predictable script; others genuinely need a flexible agent. Picking wrong costs money or reliability.

Scoping a project with SMART

Next week is a hackathon, so your idea has to fit the clock. A vague goal like "build an AI helper" sinks projects. The SMART checklist turns a wish into a plan you can actually finish.

• Specific — one clear thing. "A study-buddy that quizzes me on my biology notes," not "an education AI."
• Measurable — how do you know it worked? "Generates 5 correct quiz questions from a pasted page."
• Achievable — buildable with your tools and time (an afternoon, free APIs, no giant dataset to collect).
• Relevant — it matters to you or your audience, so you'll actually push it to the finish.
• Time-bound — a demo-able version by the end of the hackathon, not "someday."

Part 2 · Practical (90 min)

Build the "DataExplainer" agent

Now put an agent on top of structured data. The DataExplainer takes a plain-English question about the Titanic dataset, picks the right tool, and answers with a number or a chart. You'll give it three tools — the full, runnable version is in the Week 3 notebook.

Step 1 — Load the data & describe it

Load the Titanic table with pandas, then give the agent a describe_dataset() tool so it can learn the columns before it reasons.

import pandas as pd
df = pd.read_csv("https://.../titanic.csv")

def describe_dataset() -> str:
    """Tell the agent what columns and types exist."""
    return f"Rows: {len(df)}\nColumns:\n{df.dtypes}"

Step 2 — Query the data

Give it a run_query() tool that filters and aggregates with pandas .query() and .groupby(). This is the tool it reaches for to answer "how many?" or "what's the average?"

def run_query(expr: str) -> str:
    """Filter/aggregate the dataframe and return the result."""
    result = df.query(expr)          # e.g. "Sex == 'female' and Pclass == 1"
    return f"{len(result)} rows · survival rate {result['Survived'].mean():.0%}"

Step 3 — Plot a chart

Finally a plot_chart() tool using matplotlib/seaborn, so a question can be answered with a picture, not just a number.

import seaborn as sns, matplotlib.pyplot as plt

def plot_chart(kind: str, x: str, y: str) -> str:
    """Draw a chart, e.g. kind='bar', x='Pclass', y='Survived'."""
    sns.barplot(data=df, x=x, y=y)
    plt.savefig("chart.png")
    return "chart saved"

⚡ Energizer · Script or Agent? The Sorting Showdown

Before the build, spend 10–14 minutes making the deterministic-vs-agentic call fast and physical — teams race to classify real tasks and defend their choice on cost, reliability, and flexibility.

Homework & deliverables

• Finalize your project charter by the end of the session — name & objective, 2–3 core tools, data sources, and expected outputs. This is your Week 4 hackathon blueprint.
• Sanity-check it: read your one-sentence objective and confirm it hits all five SMART letters. Trim anything that makes it too big to demo.
• Line up your data — find the dataset or API you'll use (Kaggle is perfect) so you're not hunting for it next week.

Recap. A script is cheap, predictable, and fixed; an agent is flexible but pricier and less predictable — so use the simplest thing that works. You built a DataExplainer with three tools (describe, query, chart) over the Titanic data, and scoped your own project with SMART into a finished charter. Next week: you build it for real.