Building Data Apps

Use this skill to create interactive web applications that let stakeholders explore data, interact with ML models, and access analytics without writing code.

When to use this skill

Stakeholder dashboards — executives, product managers, or clients need self-service data access
ML model demos — let users test predictions with their own inputs
Internal data tools — operations teams need forms, filters, and reporting
Data exploration for non-coders — business users need to drill into datasets
Prototyping before full engineering — validate UX quickly with Python
A/B testing interfaces — experiment with different presentations of results
Multi-user analytics — shared tools accessed via browser (not notebooks)

When NOT to use this skill

Use a different skill for these related but distinct tasks:

Instead of...	Use this skill	Because...
Creating reproducible analysis notebooks	`working-in-notebooks`	Notebooks are for analysts; apps are for stakeholders
Jupyter-style interactive exploration	`working-in-notebooks`	Use Jupyter/marimo when the user is writing code
Exploratory data analysis patterns	`analyzing-data`	EDA methodology (profiling, statistical tests) belongs there
Choosing visualization libraries	`analyzing-data`	Chart types and library deep-dives are covered there
Production ML feature engineering	`engineering-ml-features`	Feature engineering logic is domain-specific
Model evaluation and cross-validation	`evaluating-ml-models`	Model comparison and metrics belong there

Quick boundary check

Data app = deployed web interface with widgets, accessed via URL, used by non-coders
Notebook = code cells + outputs, run interactively by people who write code
If the user mentions "dashboard," "app," "users clicking buttons," or "share with stakeholders" → use this skill
If the user mentions "notebook," "Jupyter," "analysis," or "explore data interactively" → use working-in-notebooks

Tool selection guide

Quick decision checklist

Question	If yes, consider
Need the simplest possible API?	Streamlit
Need ML model sharing with built-in hosting?	Gradio
Need complex reactive dashboards with flexible layouts?	Panel
Need production-grade control + React ecosystem?	Dash
Need native-like UI with async support?	NiceGUI
Deploying to Hugging Face Spaces?	Gradio or Streamlit
Already using HoloViz ecosystem (hvPlot, HoloViews)?	Panel
Need desktop + web from same codebase?	NiceGUI

Framework comparison

Framework	Best For	Key Strength	Deployment
Streamlit	Rapid prototyping, ML demos	Simplest API, largest community	Streamlit Cloud, Docker
Panel	Complex dashboards, reactive UIs	Flexible layouts, Jupyter integration	Panel serve, Cloud Run
Gradio	ML model demos, quick sharing	Built-in sharing, Hugging Face integration	Spaces, self-hosted
Dash	Production dashboards, fine control	React backend, extensive components	Gunicorn, cloud platforms
NiceGUI	Desktop + web apps, async workflows	Native-like UI, modern Python async	Native, Docker, cloud

Core workflow: Building a data app

Step 1: Choose your framework

See the decision checklist above. For most ML demos and simple dashboards → Streamlit. For complex reactive layouts → Panel. For quick ML model sharing → Gradio.

Step 2: Set up the project structure

my-app/
├── app.py              # Main entry point
├── requirements.txt    # Dependencies
├── .env               # Environment variables (not committed)
├── data/              # Data files
└── utils/             # Helper modules

Step 3: Build the minimum viable app

Start with one widget and one output. Test with real users before adding complexity.

Step 4: Handle secrets properly

# ✅ Use environment variables or framework secrets
import os
api_key = os.environ.get("OPENAI_API_KEY")

# Streamlit specific:
# api_key = st.secrets["openai_api_key"]

# ❌ Never hardcode secrets
# api_key = "sk-abc123..."

Step 5: Add caching for performance

See framework-specific advanced references for caching patterns.

Step 6: Deploy

Choose based on your needs: free tier (Streamlit Cloud, Hugging Face Spaces), containerized (Docker), or enterprise cloud (AWS, GCP, Azure).

Quick start: Streamlit

# app.py
import streamlit as st
import pandas as pd
import plotly.express as px

st.title("Sales Dashboard")

# Sidebar controls
region = st.sidebar.selectbox("Region", ["All", "North", "South", "East", "West"])

# Load data (use caching in production)
df = pd.read_parquet("sales.parquet")
if region != "All":
    df = df[df['region'] == region]

# Metrics row
col1, col2, col3 = st.columns(3)
col1.metric("Total Sales", f"${df['sales'].sum():,.0f}")
col2.metric("Orders", len(df))
col3.metric("Avg Order", f"${df['sales'].mean():.2f}")

# Visualization
fig = px.line(df.groupby('date')['sales'].sum().reset_index(), x='date', y='sales')
st.plotly_chart(fig, use_container_width=True)

# Data table
st.dataframe(df.head(100))

Run: streamlit run app.py

Quick start: Gradio

import gradio as gr
from transformers import pipeline

# Load model (example: sentiment analysis)
classifier = pipeline("sentiment-analysis")

def predict(text):
    result = classifier(text)[0]
    return result['label'], result['score']

interface = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=2, placeholder="Enter text..."),
    outputs=[gr.Label(label="Sentiment"), gr.Number(label="Confidence")],
    title="Sentiment Analysis",
    description="Enter text to analyze sentiment",
    examples=["I love this!", "This is terrible."]
)

interface.launch()

Quick start: Panel

import panel as pn
import hvplot.pandas
import pandas as pd

pn.extension()

df = pd.read_parquet("data.parquet")

# Widgets
region = pn.widgets.Select(name='Region', options=['All'] + df['region'].unique().tolist())
metric = pn.widgets.RadioBoxGroup(name='Metric', options=['sales', 'profit', 'units'])

# Reactive function
@pn.depends(region, metric)
def plot(region, metric):
    data = df if region == 'All' else df[df['region'] == region]
    return data.hvplot.line(x='date', y=metric, title=f'{metric.title()} by Date')

# Layout
app = pn.Column(
    "# Sales Dashboard",
    pn.Row(region, metric),
    plot
)

app.servable()

Run: panel serve app.py --show

Quick start: Dash

from dash import Dash, html, dcc, callback, Output, Input
import plotly.express as px

app = Dash(__name__)

df = px.data.gapminder()

app.layout = html.Div([
    html.H1("Dashboard"),
    dcc.Dropdown(df.country.unique(), 'Canada', id='country'),
    dcc.Graph(id='graph')
])

@callback(
    Output('graph', 'figure'),
    Input('country', 'value')
)
def update_graph(country):
    return px.line(df[df.country == country], x='year', y='pop')

if __name__ == '__main__':
    app.run(debug=True)

Quick start: NiceGUI

from nicegui import ui
import pandas as pd

df = pd.read_parquet("data.parquet")

# Simple UI with async support
async def load_data():
    ui.notify(f"Loaded {len(df)} rows")
    table.update_rows(df.head(20).to_dict('records'))

ui.label('Sales Dashboard').classes('text-2xl font-bold')

ui.button('Load Data', on_click=load_data)

table = ui.table(
    columns=[{'name': col, 'label': col} for col in df.columns],
    rows=[]
).classes('w-full')

ui.run()

Core design principles

1) Start simple, iterate

MVP with one widget + one visualization
Add complexity only when needed
Test with real users early

2) Optimize for the audience

Audience	Approach
Executives	Key metrics, simple filters, clean layout
Data scientists	Raw data access, parameter tuning, debug info
Operations	Refresh buttons, alerts, mobile-friendly

3) Handle state carefully

Each framework has different state management:

# Streamlit: session_state for persistence
if 'counter' not in st.session_state:
    st.session_state.counter = 0

# Dash: dcc.Store component
# Panel: param.Parameter with @depends
# NiceGUI: reactive variables or app.storage

4) Never expose secrets

# ✅ Use environment variables
import os
api_key = os.environ.get("OPENAI_API_KEY")

# Streamlit Cloud specific:
# api_key = st.secrets["openai_api_key"]

# ❌ Never hardcode
# api_key = "sk-..."

Validation and feedback loop

Self-check questions

Before deploying:

Can a non-technical user understand the interface?
Are secrets loaded from environment variables?
Is data cached appropriately (no repeated loading)?
Are there loading states for slow operations?
Does it handle edge cases (empty data, errors)?
Is it responsive on mobile if needed?
Are dependencies pinned in requirements.txt?

Performance checklist

Data loading is cached or happens once
Long computations don't block the UI
Large datasets use pagination or sampling
Images/assets are optimized

Progressive disclosure

Core references (in this skill)

references/streamlit-advanced.md — Caching, multipage apps, secrets, custom components
references/panel-advanced.md — Parameterized classes, reactive programming, layouts
references/gradio-advanced.md — Interface types, custom components, Hugging Face Spaces
references/dash-advanced.md — Callback patterns, state management, production deployment
references/nicegui-guide.md — Core concepts, async patterns, desktop vs web
references/framework-selection.md — Detailed comparison, decision framework, migration paths
references/deployment-patterns.md — Streamlit Cloud, Hugging Face, Docker, cloud platforms

Related skills

Skill	Relationship	When to use
`working-in-notebooks`	Distinct boundary	Creating analysis notebooks for coders — not stakeholder apps
`analyzing-data`	Complementary	EDA patterns, visualization library selection
`evaluating-ml-models`	Complementary	Model metrics and comparison for app display
`engineering-ml-features`	Complementary	Feature engineering behind app predictions

Common anti-patterns

❌ Loading data on every interaction (use caching)
❌ Blocking the UI with long computations (use async or progress indicators)
❌ No error handling for edge cases (empty data, network failures)
❌ Hardcoded file paths or credentials (use environment variables)
❌ Too many widgets (cognitive overload — prioritize)
❌ No mobile consideration when audience uses phones
❌ Confusing notebook code with app code (different paradigms)

building-data-apps