app.py

import streamlit as st
import pandas as pd
import plotly.express as px
import plotly.graph_objects as go
import plotly.figure_factory as ff
from plotly.subplots import make_subplots
import json
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
import re
import numpy as np


st.set_page_config(
    page_title="The Language of Rejection",
    page_icon="💔",
    layout="wide",
    initial_sidebar_state="expanded"
)

@st.cache_data
def load_data():
    df = pd.read_csv('data/rejection_analysis_extended.csv')
    df_clean = df[df['status'] != 'ghosted'].copy()
    
    with open('data/shap_results_all.json', 'r') as f:
        shap_data = json.load(f)
    
    return df_clean, shap_data

df, shap_results = load_data()
vader = SentimentIntensityAnalyzer()


@st.cache_data
def process_all_shap():
    """Extract all words with their impacts across companies"""
    all_negative = []
    all_positive = []
    
    for company, data in shap_results.items():
        for word, score in data['words']:
            if score < 0:
                all_negative.append({
                    'word': word,
                    'score': score,
                    'company': company,
                    'vader': data['vader'],
                    'roberta': data['roberta']
                })
            else:
                all_positive.append({
                    'word': word,
                    'score': score,
                    'company': company,
                    'vader': data['vader'],
                    'roberta': data['roberta']
                })
    
    df_neg = pd.DataFrame(all_negative)
    df_pos = pd.DataFrame(all_positive)
    
    return df_neg, df_pos

df_negative, df_positive = process_all_shap()


def analyze_text(text):
    """Analyze email text"""
    score = vader.polarity_scores(text)['compound']
    words = text.lower().split()
    
    joy_keywords = ['hope', 'happy', 'good', 'luck', 'best', 'wish', 'encourage']
    apology_keywords = ['sorry', 'unfortunately', 'regret', 'apologies', 'apologize']
    positive_keywords = ['thank', 'appreciate', 'grateful', 'impressed', 'value', 
                        'strong', 'excellent', 'great', 'pleased', 'interested']
    
    joy_count = sum(1 for w in words if any(j in w for j in joy_keywords))
    apology_count = sum(1 for w in words if any(a in w for a in apology_keywords))
    positive_count = sum(1 for w in words if any(p in w for p in positive_keywords))
    
    return {
        'score': score,
        'joy': joy_count,
        'apology': apology_count,
        'positive': positive_count,
        'word_count': len(words)
    }


st.sidebar.title("💔 Navigation")
if "page" not in st.session_state:
    st.session_state.page = "🏠 The Story"

pages = ["🏠 The Story", "📊 The Data", "🔬 Deep Dive: SHAP", "🧪 Try It Yourself"]

for page in pages:
    if st.session_state.page == page:
        st.sidebar.button(
            f"✅ {page}",
            use_container_width=True,
            disabled=True,
            type="primary"
        )
    else:
        if st.sidebar.button(page, use_container_width=True):
            st.session_state.page = page
            st.rerun()

st.sidebar.markdown("---")
st.sidebar.markdown("### 📌 Quick Stats")
st.sidebar.metric("Emails Analyzed", "14")
st.sidebar.metric("Joy Factor", "+0.605")
st.sidebar.metric("The Magic Ratio", "4:1")

st.sidebar.markdown("---")
st.sidebar.markdown("[GitHub](https://github.com/jgchoti/job_rejection_analysis) | [LinkedIn](https://www.linkedin.com/in/chotirat/)")

if st.session_state.page == "🏠 The Story":
    st.title("💔 The Language of Rejection")
    st.markdown("### What makes some rejection emails feel warm while others feel crushing?")
    
    st.markdown("---")
    
    col1, col2, col3 = st.columns([1, 2, 1])  
    with col2:
        st.image("assets/ophelia.jpg", caption="*Ophelia* (1851–52) — Sir John Everett Millais",)
    
    # Introduction
    st.markdown("""
    <div style="background-color: #FEF3E5; padding: 25px; border-radius: 12px; border-left: 6px solid #CC7C60;">
    <h2>📖 How This Started</h2>
    <p style="font-size: 1.1rem; line-height: 1.8;">
    After receiving many job rejection, something clicked. Instead of feeling frustrated, 
    I got curious: <b>could I quantify what makes a rejection feel "warm" or "cold"?</b>
    </p>
    <p style="font-size: 1.1rem; line-height: 1.8;">
    I collected all rejection emails (2024-2025), anonymized them, and analyzed them using 
    Natural Language Processing(NLP).
    </p>
    <p style="font-size: 1rem; margin-top: 10px;">
    For the full code, visit my <a href="https://github.com/jgchoti/job_rejection_analysis" target="_blank" >GitHub repository</a>.
    </p>
    </div>
    """, unsafe_allow_html=True)
    st.markdown("---")
    st.markdown("")
    st.markdown("### 📧 The Tale of Two Emails")
    
    col1, col2 = st.columns(2)
    
    with col1:
        st.markdown("""
        <div style="background-color: #ffebee; padding: 20px; border-radius: 10px; border-left: 5px solid #e53935;">
        <h4 style="color: #c62828; margin-top: 0;">❌ The Coldest Email (0.307)</h4>
        <ul style="font-size: 1rem; line-height: 1.6;">
            <li>20 words total</li>
            <li>2 apologies</li>
            <li>2 positive words</li>
            <li>0 joy words</li>
        </ul>
        <p style="font-style: italic; color: #666; margin-bottom: 0;">
        "Thanks... Unfortunately... We're sorry..."
        </p>
        </div>
        """, unsafe_allow_html=True)
    
    with col2:
        st.markdown("""
        <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px; border-left: 5px solid #43a047;">
        <h4 style="color: #2e7d32; margin-top: 0;">✅ The Warmest Email (0.990)</h4>
        <ul style="font-size: 1rem; line-height: 1.6;">
            <li>88 words total</li>
            <li>0 apologies</li>
            <li>17 positive words</li>
            <li>2 joy words</li>
        </ul>
        <p style="font-style: italic; color: #666; margin-bottom: 0;">
        "We were impressed... We wish you the best..."
        </p>
        </div>
        """, unsafe_allow_html=True)
    st.write("") 
    st.warning("""***The difference?*** 68 words.
    Completely different emotional impact.
    """)
    
    # Key Findings
    st.markdown("### 🔍 What I Discovered")
    
    finding_col1, finding_col2, finding_col3 = st.columns(3)
    
    with finding_col1:
        st.markdown("""
        <div style="background-color: white; padding: 20px; border-radius: 10px; box-shadow: 0 2px 8px rgba(0,0,0,0.1); height: 100%;">
        <h3 style="color: #3498db; margin-top: 0;">🎉 Finding #1</h3>
        <h4>Joy Words Win</h4>
        <p style="font-size: 0.95rem; line-height: 1.6;">
        Words like "hope," "wish," and "best" correlate <b>+0.605</b> with warmth.
        That's 3× stronger than any other factor.
        </p>
        <p style="font-size: 0.9rem; color: #7f8c8d; margin-bottom: 0;">
        <i>The warmest emails average 2 joy words. The coldest? Zero.</i>
        </p>
        </div>
        """, unsafe_allow_html=True)
    
    with finding_col2:
        st.markdown("""
        <div style="background-color: white; padding: 20px; border-radius: 10px; box-shadow: 0 2px 8px rgba(0,0,0,0.1); height: 100%;">
        <h3 style="color: #e74c3c; margin-top: 0;">⚖️ Finding #2</h3>
        <h4>The 4:1 Rule</h4>
        <p style="font-size: 0.95rem; line-height: 1.6;">
        Each apology ("sorry," "unfortunately") needs <b>at least 4 positive words</b> 
        to maintain warmth.
        </p>
        <p style="font-size: 0.9rem; color: #7f8c8d; margin-bottom: 0;">
        <i>Below 4:1 = guaranteed cold. Above 6:1 = safe zone.</i>
        </p>
        </div>
        """, unsafe_allow_html=True)
    
    with finding_col3:
        st.markdown("""
        <div style="background-color: white; padding: 20px; border-radius: 10px; box-shadow: 0 2px 8px rgba(0,0,0,0.1); height: 100%;">
        <h3 style="color: #9b59b6; margin-top: 0;">🧪 Finding #3</h3>
        <h4>Context Matters</h4>
        <p style="font-size: 0.95rem; line-height: 1.6;">
        Transformers detect <b>empty politeness </b> that simple word-counting overlook.
        </p>
        <p style="font-size: 0.9rem; color: #7f8c8d; margin-bottom: 0;">
        <i>Some emails score 0.97 (lexicon) but -0.17 (transformer).</i>
        </p>
        </div>
        """, unsafe_allow_html=True)
    
    st.markdown("---")
    
    # Why It Matters
    st.markdown("""
    <div style="background-color: #FEF3E5; padding: 25px; border-radius: 12px; border-left: 6px solid #CC7C60;">
    <h2>💡 Why This Matters</h2>
    <div style="font-size: 1.05rem; line-height: 1.8; color: #212121;">
    <p><b>For job seekers like me:</b> That sinking feeling after a cold rejection? 
    It's not personal. It's just a bad template. Some companies accidentally crush spirits. 
    Others lift them. Neither is intentional.</p>

    <p><b>For business:</b> Small improvements to templates—like warm language and clear next steps—aren’t just nice. With AI reading emails first, setting the right tone helps algorithms recognize your intent as positive, influencing how messages are categorized and analyzed. Same automation. Different outcomes.</p>

    <p style="margin-bottom: 0;"><b>The bigger point:</b> Rejection is unavoidable in hiring. 
    Cruelty is optional.</p>
    </div>
    </div>
    """, unsafe_allow_html=True)

    
    st.markdown("---")
    st.markdown("👉 **Explore the data in the next section to see exactly how I measured this.**")
    if st.button("Next → 📊 The Data"):
        st.session_state.page = "📊 The Data"
        st.rerun()
elif st.session_state.page == "📊 The Data":
    st.title("📊 The Data Behind the Story")
    st.markdown("### Let's look at the numbers that prove these patterns are real")
    
    st.markdown("---")
    
    # Overview metrics
    st.markdown("## 📈 Dataset Overview")
    
    metric_col1, metric_col2, metric_col3, metric_col4 = st.columns(4)
    
    with metric_col1:
        st.metric("Total Emails", "14", help="Real rejection emails from 2024-2025 job search")
    with metric_col2:
        st.metric("Warmest Score", "0.990", delta="+0.68 vs coldest")
    with metric_col3:
        st.metric("Coldest Score", "0.307", delta="-0.68 vs warmest", delta_color="inverse")
    with metric_col4:
        st.metric("Average Score", f"{df['vader_compound'].mean():.3f}")
    
    st.markdown("---")
    
    # Finding #1: Joy Factor
    st.markdown("## 🎉 Finding #1: The Joy Factor")
    st.markdown("""
    **The Question:** Which emotions predict warmth best?
    
    **The Answer:** Joy-based words correlate **+0.605** with warmth. That's:
    - 5× stronger than trust (+0.116)
    - 10× stronger than anticipation (+0.057)
    """)

        
    with st.expander("### 🔥 Emotion-Sentiment Heatmap", expanded=False):
        st.markdown("### 🔥 Emotion-Sentiment Heatmap: Which Models Respond to Which Emotions?")
        st.markdown("""
        This heatmap reveals a critical insight: **lexicon-based models** (VADER, AFINN, TextBlob) 
        respond strongly to joy, while **transformers** (RoBERTa, SST-2) are less influenced by simple emotion counts.
        """)

        emotion_cols = ['emotion_joy', 'emotion_trust', 'emotion_anticipation', 'emotion_sadness', 'emotion_fear', 'emotion_anger']
        emotion_names = ['Joy', 'Trust', 'Anticipation', 'Sadness', 'Fear', 'Anger']
        model_cols = ['vader_compound', 'hf_roberta_score', 'textblob_polarity','afinn_score', 'hf_sst2_score']
        model_names = ['VADER', 'RoBERTa', 'TextBlob', 'AFINN', 'SST-2']

        # Calculate correlation matrix (emotions × models)
        corr_matrix = np.zeros((len(emotion_cols), len(model_cols)))

        for i, emotion in enumerate(emotion_cols):
            for j, model in enumerate(model_cols):
                corr_matrix[i, j] = df[emotion].corr(df[model])

        # Create heatmap
        fig3 = go.Figure(data=go.Heatmap(
        z=corr_matrix,
        x=model_names,
        y=emotion_names,
        colorscale='RdYlGn',
        zmid=0,
        text=np.round(corr_matrix, 3),
        texttemplate='%{text}',
        textfont={"size": 13, "color": "black"},
        colorbar=dict(
        title='Correlation<br>Strength',
        thickness=20,
        len=0.7
        ),
        hovertemplate='<b>%{y} → %{x}</b><br>Correlation: %{z:.3f}<extra></extra>'
))

# Highlight Joy-VADER (strongest)
        fig3.add_shape(
    type='rect',
    x0=-0.5, x1=0.5,
    y0=-0.5, y1=0.5,
    line=dict(color='#2ecc71', width=4)
)

# Highlight Joy row
        fig3.add_shape(
    type='rect',
    x0=-0.5, x1=4.5,
    y0=-0.5, y1=0.5,
    line=dict(color='#3498db', width=2, dash='dash')
)

        fig3.update_layout(
    title='How Each Model Responds to Different Emotions',
    xaxis_title='Sentiment Model',
    yaxis_title='Emotion Type',
    template='plotly_white',
    height=550,
    xaxis={
        'side': 'bottom',
        'tickangle': 0
    },
    yaxis={
        'autorange': 'reversed'  
    },
    )

        st.plotly_chart(fig3, use_container_width=True)

        col1, col2 = st.columns(2)

        with col1:
            st.markdown("""
        <div style="background-color: #e8f5e9; padding: 15px; border-radius: 10px; border-left: 4px solid #2ecc71;">
        <h4 style="color: #1e8449; margin-top: 0;">🤖 Lexicon Models (VADER, AFINN, TextBlob)</h4>
        <ul style="font-size: 0.95rem;">
        <li><b>Joy: +0.605</b> (very strong)</li>
        <li><b>Trust: +0.116</b> (weak)</li>
        <li><b>Anticipation: +0.057</b> (very weak)</li>
        </ul>
        <p style="margin-bottom: 0; font-size: 0.9rem;"><i>These models COUNT words, so joy words = high scores</i></p>
        </div>
        """, unsafe_allow_html=True)

        with col2:
            st.markdown("""
        <div style="background-color: #eef2f7; padding: 15px; border-radius: 10px; border-left: 4px solid #3498db;">
        <h4 style="color: #2874a6; margin-top: 0;">🧠 Transformer Models (RoBERTa, SST-2)</h4>
        <ul style="font-size: 0.95rem;">
        <li><b>Joy: +0.38</b> (moderate)</li>
        <li><b>Less emotion-driven</b></li>
        <li><b>Focus on context/phrases</b></li>
        </ul>
        <p style="margin-bottom: 0; font-size: 0.9rem;"><i>These models understand rejection PATTERNS, not just word counts</i></p>
        </div>
        """, unsafe_allow_html=True)
        
        st.markdown("")

    fig1 = px.scatter(
        df,
        x='emotion_joy',
        y='vader_compound',
        size='afinn_positive_count',
        color='vader_compound',
        hover_data=['company_id', 'apology_words'],
        title="Joy Words Strongly Predict Warmth",
        labels={
            'emotion_joy': 'Number of Joy Words (hope, wish, best, good luck, etc.)',
            'vader_compound': 'Warmth Score',
            'afinn_positive_count': 'Total Positive Words'
        },
        color_continuous_scale='RdYlGn',
        height=500
    )
    fig1.update_traces(marker=dict(line=dict(width=1.5, color='white')))
    fig1.update_layout(
        template='plotly_white',
        font=dict(size=12),
        title_font_size=16
    )
    st.plotly_chart(fig1, use_container_width=True)
    
    st.info("**💡 Key Insight:** Emails with 2+ joy words score 0.95+. Emails with 0 joy words score below 0.50. The pattern is clear and consistent.")

    st.markdown("### 📊 All Emotions Compared")
    st.markdown("Joy isn't the only emotion - let's see how all emotions correlate with warmth:")

    corr_with_sentiment = []
    for col in emotion_cols:
        corr = df[col].corr(df['vader_compound'])
        corr_with_sentiment.append(corr)
    fig2 = go.Figure()

    colors = ['#2ecc71' if c > 0.2 else '#95a5a6' for c in corr_with_sentiment]

    fig2.add_trace(go.Bar(
    x=emotion_names,
    y=corr_with_sentiment,
    marker_color=colors,
    text=[f'{c:+.3f}' for c in corr_with_sentiment],
    textposition='outside',
    hovertemplate='<b>%{x}</b><br>Correlation: %{y:.3f}<extra></extra>'
    ))

    # Highlight joy
    fig2.add_annotation(
    x='Joy',
    y=corr_with_sentiment[0] + 0.1,
    text='<b>5× stronger than<br>next best!</b>',
    showarrow=True,
    arrowhead=2,
    arrowcolor='#2ecc71',
    font=dict(color='#2ecc71', size=13),
    bgcolor='#e8f5e9',
    bordercolor='#2ecc71',
    borderwidth=2
    )

    fig2.update_layout(
    title='Emotion Correlation with Warmth Score',
    xaxis_title='Emotion Type',
    yaxis_title='Correlation with Warmth (r)',
    template='plotly_white',
    height=450,
    showlegend=False
    )

    fig2.add_hline(y=0, line_color='gray', line_width=1)

    st.plotly_chart(fig2, use_container_width=True)

    st.info("""
    **📊 The Verdict:** 
    Joy dominates with **+0.605 correlation** - that's:
    - 5× stronger than trust (+0.116)
    - 10× stronger than anticipation (+0.057)
    - **3× stronger than ALL other emotions combined**

    This isn't just one of many factors. Joy words are THE factor.
    """)
  

    st.markdown("---")
    
    # Finding #2: The 4:1 Rule
    st.markdown("## ⚖️ Finding #2: The 4:1 Compensation Rule")
    st.markdown("""
    **The Question:** How many positive words are needed to balance one apology?

    **The Answer:** At least **4 positive words per apology**. But the best strategy? **Zero apologies.**
    """)

    # Separate companies with/without apologies
    df_with_apologies = df[df['apology_words'] > 0].copy()
    df_no_apologies = df[df['apology_words'] == 0].copy()

    # Calculate ratio for those with apologies
    df_with_apologies['ratio'] = df_with_apologies['afinn_positive_count'] / df_with_apologies['apology_words']
    df_with_apologies['zone'] = pd.cut(
        df_with_apologies['ratio'], 
        bins=[0, 3, 6, 20], 
        labels=['❌ Danger (<4:1)', '⚠️ Minimum (4-6:1)', '✅ Safe (6:1+)']
    )

    # Show info box about zero-apology companies
    if len(df_no_apologies) > 0:
        companies_list = ', '.join(df_no_apologies['company_id'].tolist())
        avg_score = df_no_apologies['vader_compound'].mean()
    
        st.info(f"""
        **🏆 The Ultimate Strategy: {len(df_no_apologies)} companies avoided apologies entirely:**
    
        **{companies_list}**
    
        - Average warmth score: **{avg_score:.3f}** (highest group!)
        - No compensation needed - they never triggered the 4:1 rule
        - This is the gold standard ✨
    
        The analysis below focuses on companies that DID apologize and whether they compensated enough.
        """)

    # Create scatter plot for companies WITH apologies
    fig3 = px.scatter(
        df_with_apologies,
        x='ratio',
        y='vader_compound',
        color='zone',
        text='company_id',
        title="The 4:1 Rule: Positive-to-Apology Ratio vs Warmth (For Companies That Apologized)",
        labels={
        'ratio': 'Positives per Apology',
        'vader_compound': 'Warmth Score'
        },
        color_discrete_map={
        '❌ Danger (<4:1)': '#e74c3c',
        '⚠️ Minimum (4-6:1)': '#f39c12',
        '✅ Safe (6:1+)': '#2ecc71'
        },
        height=500
    )

    fig3.update_traces(textfont=dict(size=8), textposition='bottom center', marker=dict(size=15, line=dict(width=2, color='white')))
    fig3.add_vline(x=4, line_dash="dash", line_color="red", line_width=2,
              annotation_text="Critical threshold")
    fig3.add_vline(x=6, line_dash="dash", line_color="green", line_width=2,
              annotation_text="Safe zone starts")
    fig3.add_hline(y=0.85, line_dash="dash", line_color="gray",
              annotation_text="Warm threshold (0.85)")

    # Add annotation about zero-apology companies
    if len(df_no_apologies) > 0:
        fig3.add_annotation(
        x=0.98, y=1.2,
        xref='paper', yref='paper',
        text=f'<b>{len(df_no_apologies)} companies with 0 apologies<br>scored {avg_score:.3f} average<br>(not shown - off the chart!)</b>',
        showarrow=False,
        bgcolor='#e3f2fd',
        bordercolor='#2196f3',
        borderwidth=2,
        font=dict(size=11),
        align='left',
        xanchor='right',
        yanchor='top'
    )

    fig3.update_layout(template='plotly_white')
    st.plotly_chart(fig3, use_container_width=True)

    # Zone statistics
    st.markdown("### 📊 Success Rate by Zone (Among Companies That Apologized)")
    col1, col2, col3 = st.columns(3)

    danger_zone = df_with_apologies[df_with_apologies['zone'] == '❌ Danger (<4:1)']
    minimum_zone = df_with_apologies[df_with_apologies['zone'] == '⚠️ Minimum (4-6:1)']
    safe_zone = df_with_apologies[df_with_apologies['zone'] == '✅ Safe (6:1+)']

    with col1:
        st.markdown("""
        <div style="background-color: #ffebee; padding: 20px; border-radius: 10px;">
        <h4 style="color: #c62828;">❌ Danger Zone</h4>
        <p style="font-size: 1.8rem; font-weight: bold; margin: 10px 0;">0%</p>
        <p style="margin-bottom: 0;">None are warm (0.85+)</p>
        </div>
        """, unsafe_allow_html=True)

    with col2:
        warm_min = (minimum_zone['vader_compound'] >= 0.85).sum()
        total_min = len(minimum_zone)
        pct_min = (warm_min / total_min * 100) if total_min > 0 else 0
        st.markdown(f"""
        <div style="background-color: #fff3e0; padding: 20px; border-radius: 10px;">
        <h4 style="color: #f57c00;">⚠️ Minimum Zone</h4>
        <p style="font-size: 1.8rem; font-weight: bold; margin: 10px 0;">{pct_min:.0f}%</p>
        <p style="margin-bottom: 0;">{warm_min} of {total_min} are warm</p>
        </div>
        """, unsafe_allow_html=True)

    with col3:
        warm_safe = (safe_zone['vader_compound'] >= 0.85).sum()
        total_safe = len(safe_zone)
        pct_safe = (warm_safe / total_safe * 100) if total_safe > 0 else 0
        st.markdown(f"""
        <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px;">
        <h4 style="color: #2e7d32;">✅ Safe Zone</h4>
        <p style="font-size: 1.8rem; font-weight: bold; margin: 10px 0;">{pct_safe:.0f}%</p>
        <p style="margin-bottom: 0;">{warm_safe} of {total_safe} are warm</p>
        </div>
        """, unsafe_allow_html=True)

    st.markdown("")
    st.success("""
    **💡 Key Insight:** 
    - Below 4:1 ratio = 0% success rate
    - Above 4:1 = guaranteed warm
    - **But the real winners?** The companies that never apologized at all! 🏆
    """)

    st.markdown("---")

    
    # All companies ranked
    st.markdown("## 📊 All 14 Companies Ranked")
    st.markdown("From warmest to coldest rejection emails:")
    
    df_sorted = df.sort_values('vader_compound', ascending=False)
    
    fig4 = go.Figure()
    colors = ['#2ecc71' if score >= 0.95 else '#3498db' if score >= 0.85 else '#f39c12' if score >= 0.60 else '#e74c3c' 
              for score in df_sorted['vader_compound']]
    
    fig4.add_trace(go.Bar(
        x=df_sorted['company_id'],
        y=df_sorted['vader_compound'],
        marker_color=colors,
        text=[f'{v:.3f}' for v in df_sorted['vader_compound']],
        textposition='outside',
        hovertemplate='<b>%{x}</b><br>Score: %{y:.3f}<br><extra></extra>'
    ))
    
    fig4.add_hline(y=0.95, line_dash="dash", line_color="green", annotation_text="Very Warm (0.95+)")
    fig4.add_hline(y=0.85, line_dash="dash", line_color="orange", annotation_text="Warm (0.85+)")
    fig4.add_hline(y=0.60, line_dash="dash", line_color="gray", annotation_text="Neutral (0.60+)")
    
    fig4.update_layout(
        title="Company Warmth Rankings",
        xaxis_title="Company",
        yaxis_title="Warmth Score (VADER)",
        template='plotly_white',
        height=500,
        showlegend=False,
        xaxis={'tickangle': -45}
    )
    
    st.plotly_chart(fig4, use_container_width=True)
    
    # Summary table
    st.markdown("### 📋 Detailed Breakdown")
    summary_df = df_sorted[['company_id', 'vader_compound', 'emotion_joy', 'afinn_positive_count', 'apology_words']].copy()
    summary_df.columns = ['Company', 'Warmth Score', 'Joy Words', 'Positive Words', 'Apologies']
    summary_df = summary_df.round(3)
    st.dataframe(summary_df, use_container_width=True, hide_index=True)
    

elif st.session_state.page == "🔬 Deep Dive: SHAP":
    st.title("🔬 Deep Dive: SHAP Word Analysis")
    st.markdown("### Which specific words push sentiment up or down? Let's find out.")
    
    st.markdown("""
    **SHAP (SHapley Additive exPlanations)** reveals the exact contribution of each word 
    to RoBERTa's sentiment decision. This is where I find the **real insights** - 
    which words companies should use, and which they should avoid.
    """)
    
    st.markdown("---")
    
    # Tab structure for different views
    shap_tab1, shap_tab2, shap_tab3, shap_tab4 = st.tabs([
        "☠️ Words to Avoid",
        "⭐ Words to Use", 
        "🔄 Substitution Guide",
        "📚 Case Studies"
    ])
    

    with shap_tab1:
        st.markdown("## ☠️ The Most Damaging Words")
        st.markdown("These words carry the most negative weight across all emails:")
        
        # Get top 15 most negative
        top_negative = df_negative.nsmallest(15, 'score')
        
        fig_neg = go.Figure(data=[
            go.Bar(
                x=top_negative['score'],
                y=[f"{row['word']} ({row['company']})" for _, row in top_negative.iterrows()],
                orientation='h',
                marker_color='#e74c3c',
                text=[f"{score:.3f}" for score in top_negative['score']],
                textposition='outside',
                hovertemplate='<b>%{y}</b><br>Impact: %{x:.3f}<extra></extra>'
            )
        ])
        
        fig_neg.update_layout(
            title="Top 15 Most Damaging Words (Across All Companies)",
            xaxis_title="SHAP Value (Negative Impact)",
            yaxis_title="Word (Company)",
            template='plotly_white',
            height=600,
            showlegend=False
        )
        
        st.plotly_chart(fig_neg, use_container_width=True)
        
        # Insights
        st.markdown("### 💡 Key Insights")
        
        col1, col2 = st.columns(2)
        
        with col1:
            st.markdown("""
            <div style="background-color: #F3EDFA; padding: 20px; border-radius: 10px;>
            <h4>🚨 The "Unfortunately" Disaster</h4>
            <p><b>Company_C: -0.7727</b> (worst word in dataset!)</p>
            <p>This single word crushes sentiment. VADER thinks the email is positive (0.974), 
            but RoBERTa knows it's negative (-0.168) because transformers have learned 
            "unfortunately" is a rejection pattern.</p>
            <p style="margin-bottom: 0;"><b>Impact:</b> Using "unfortunately" creates a 1.14 point 
            disagreement between lexicons and transformers.</p>
            </div>
            """, unsafe_allow_html=True)
        
        with col2:
            st.markdown("""
            <div style="background-color: #F3EDFA; padding: 20px; border-radius: 10px;>
            <h4>💀 The "Regret" Problem</h4>
            <p><b>Company_G: -0.7341</b> (second-worst!)</p>
            <p>"Regret" sounds polite, but it's actually <b>worse than "sorry"</b> (-0.47) 
            because it's strongly associated with formal rejections.</p>
            <p style="margin-bottom: 0;"><b>Advice:</b> Never use "regret" in rejection emails. 
            It's a learned rejection signal.</p>
            </div>
            """, unsafe_allow_html=True)
        
        # Summary statistics
        st.markdown("### 📊 Word Frequency Analysis")
        
        # Count how many times each negative word appears
        neg_word_counts = df_negative.groupby('word').agg({
            'score': ['mean', 'count', 'min']
        }).round(3)
        neg_word_counts.columns = ['avg_impact', 'frequency', 'worst_impact']
        neg_word_counts = neg_word_counts.sort_values('avg_impact').head(10)
        
        st.dataframe(
            neg_word_counts.reset_index().rename(columns={
                'word': 'Word',
                'avg_impact': 'Average Impact',
                'frequency': 'Appears in # Emails',
                'worst_impact': 'Worst Single Impact'
            }),
            use_container_width=True,
            hide_index=True
        )
        
        st.warning("""
        **Bottom Line:** 
        - "Unfortunately" appears in 8 emails with average impact -0.47
        - "Sorry" appears in 4 emails with average impact -0.29
        - "Regret" appears in 2 emails but has devastating impact (-0.66 average)
        
        **Recommendation:** Avoid all three. If you must acknowledge disappointment, use "disappointing" (-0.17) instead.
        """)
    

    with shap_tab2:
        st.markdown("## ⭐ The Most Helpful Words")
        st.markdown("These words consistently boost warmth:")
        
        # Get top 15 most positive
        top_positive = df_positive.nlargest(15, 'score')
        
        fig_pos = go.Figure(data=[
            go.Bar(
                x=top_positive['score'],
                y=[f"{row['word']} ({row['company']})" for _, row in top_positive.iterrows()],
                orientation='h',
                marker_color='#2ecc71',
                text=[f"+{score:.3f}" for score in top_positive['score']],
                textposition='outside',
                hovertemplate='<b>%{y}</b><br>Impact: %{x:.3f}<extra></extra>'
            )
        ])
        
        fig_pos.update_layout(
            title="Top 15 Most Helpful Words (Across All Companies)",
            xaxis_title="SHAP Value (Positive Impact)",
            yaxis_title="Word (Company)",
            template='plotly_white',
            height=600,
            showlegend=False
        )
        
        st.plotly_chart(fig_pos, use_container_width=True)
        
        # Insights
        st.markdown("### 💡 The Power Trio")
        
        col1, col2, col3 = st.columns(3)
        
        with col1:
            st.markdown("""
            <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px;">
            <h4>🏆 #1: "Appreciate"</h4>
            <p><b>Average Impact: +0.52</b></p>
            <p>Appears in 5 emails with consistently strong positive impact.</p>
            <p><b>Best Use:</b> Company_M (+0.59)</p>
            <p style="margin-bottom: 0;"><i>"We appreciate your interest..."</i></p>
            </div>
            """, unsafe_allow_html=True)
        
        with col2:
            st.markdown("""
              <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px;">
            <h4>🥈 #2: "Thank"</h4>
            <p><b>Average Impact: +0.47</b></p>
            <p>Most frequent positive word. Appears in 10+ emails.</p>
            <p><b>Best Use:</b> Company_E (+0.56)</p>
            <p style="margin-bottom: 0;"><i>"Thank you for your time..."</i></p>
            </div>
            """, unsafe_allow_html=True)
        
        with col3:
            st.markdown("""
              <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px;">
            <h4>🥉 #3: "Value"</h4>
            <p><b>Average Impact: +0.39</b></p>
            <p>Rare but powerful. Shows genuine appreciation.</p>
            <p><b>Best Use:</b> Company_F (+0.39)</p>
            <p style="margin-bottom: 0;"><i>"We value your effort..."</i></p>
            </div>
            """, unsafe_allow_html=True)
        
        # Word frequency
        st.markdown("### 📊 Positive Word Frequency")
        
        pos_word_counts = df_positive.groupby('word').agg({
            'score': ['mean', 'count', 'max']
        }).round(3)
        pos_word_counts.columns = ['avg_impact', 'frequency', 'best_impact']
        pos_word_counts = pos_word_counts.sort_values('avg_impact', ascending=False).head(10)
        
        st.dataframe(
            pos_word_counts.reset_index().rename(columns={
                'word': 'Word',
                'avg_impact': 'Average Impact',
                'frequency': 'Appears in # Emails',
                'best_impact': 'Best Single Impact'
            }),
            use_container_width=True,
            hide_index=True
        )
        
        st.success("""
        **Action Items:**
        - Use "appreciate" instead of just "thank" for stronger impact
        - Include "value" when you mean it (don't overuse)
        - "Impressed" (+0.51 peak) is powerful when specific
        - "Happy" and "wish" are joy words that boost warmth significantly
        """)
    

    with shap_tab3:
        st.markdown("## 🔄 Word Substitution Guide")
        st.markdown("### Simple swaps that make emails warmer:")
        
        # Create substitution table
        substitutions = [
            {
                'avoid': 'Sorry',
                'avoid_impact': -0.47,
                'use': 'Disappointed',
                'use_impact': -0.17,
                'improvement': '+64%',
                'example': 'We\'re disappointed we can\'t move forward'
            },
            {
                'avoid': 'Regret',
                'avoid_impact': -0.73,
                'use': 'Disappointing',
                'use_impact': -0.17,
                'improvement': '+77%',
                'example': 'This is disappointing news to share'
            },
            {
                'avoid': 'Unfortunately',
                'avoid_impact': -0.77,
                'use': '(Skip it!)',
                'use_impact': 0.0,
                'improvement': '+100%',
                'example': 'We\'ve decided to move forward with another candidate'
            },
            {
                'avoid': 'Thanks (generic)',
                'avoid_impact': -0.16,
                'use': 'Thank you + specific',
                'use_impact': +0.42,
                'improvement': '+163%',
                'example': 'Thank you for your thoughtful approach to the case study'
            },
            {
                'avoid': 'Thank (alone)',
                'avoid_impact': +0.13,
                'use': 'Appreciate',
                'use_impact': +0.54,
                'improvement': '+315%',
                'example': 'We appreciate your interest in our team'
            },
            {
                'avoid': 'Good luck',
                'avoid_impact': +0.12,
                'use': 'Wish you the best',
                'use_impact': +0.23,
                'improvement': '+92%',
                'example': 'We wish you the best in your job search'
            }
        ]
        
        df_sub = pd.DataFrame(substitutions)
        
        # Display as formatted table
        st.markdown("### 🔄 Recommended Substitutions")
        
        for _, row in df_sub.iterrows():
            col1, col2, col3 = st.columns([1, 1, 2])
            
            with col1:
                st.markdown(f"""
                <div style="background-color: #ffebee; padding: 15px; border-radius: 8px; text-align: center;">
                <h4 style="color: #c62828; margin: 0;">❌ Avoid</h4>
                <p style="font-size: 1.2rem; font-weight: bold; margin: 10px 0;">{row['avoid']}</p>
                <p style="margin: 0; color: #666;">Impact: {row['avoid_impact']:.2f}</p>
                </div>
                """, unsafe_allow_html=True)
            
            with col2:
                use_color = "#e8f5e9" if row['use_impact'] >= 0 else "#fff3e0"
                text_color = "#2e7d32" if row['use_impact'] >= 0 else "#f57c00"
                st.markdown(f"""
                <div style="background-color: {use_color}; padding: 15px; border-radius: 8px; text-align: center;">
                <h4 style="color: {text_color}; margin: 0;">✅ Use</h4>
                <p style="font-size: 1.2rem; font-weight: bold; margin: 10px 0;">{row['use']}</p>
                <p style="margin: 0; color: #666;">Impact: {row['use_impact']:+.2f}</p>
                </div>
                """, unsafe_allow_html=True)
            
            with col3:
                st.markdown(f"""
                <div style="background-color: white; padding: 15px; border-radius: 8px; border-left: 4px solid #3498db;">
                <p style="margin: 0; color: #666;"><b>Improvement:</b> {row['improvement']}</p>
                <p style="margin: 5px 0 0 0; font-style: italic;">"{row['example']}"</p>
                </div>
                """, unsafe_allow_html=True)
            
            st.markdown("<br>", unsafe_allow_html=True)
        
        # Key takeaways
        st.info("""
        **💡 Key Takeaway:** Small word changes create massive improvements:
        - Dropping "unfortunately" entirely = +100% better
        - Swapping "sorry" for "disappointed" = +64% better  
        - Using "appreciate" instead of "thank" = +315% better
        
        These are ONE-TIME template fixes that improve every rejection forever.
        """)

    with shap_tab4:
        st.markdown("## 📚 Case Studies: Winners & Losers")
        st.markdown("### Let's examine why certain companies succeeded or failed:")
        
        # Case Study 1: Company F (Winner)
        st.markdown("### 🏆 Case Study #1: Company F - The Winner (0.990)")
        
        company_f = shap_results['Company_F']
        
        col1, col2 = st.columns(2)
        
        with col1:
            st.markdown("""
            <div style="background-color: #e8f5e9; padding: 20px; border-radius: 10px;">
            <h4>✅ What They Did Right</h4>
            <ul>
                <li><b>Strong positives:</b> Used powerful words like "appreciate" (+0.54), "value" (+0.39), "happy" (+0.26)</li>
                <li><b>Smart apology choice:</b> Used "disappointing" (-0.17) instead of "sorry" (-0.47)</li>
                <li><b>Joy words:</b> Included "wish" and "best"</li>
                <li><b>No "unfortunately":</b> Avoided the worst word</li>
            </ul>
            </div>
            """, unsafe_allow_html=True)
            
            st.metric("Positive Force", f"+{company_f['words'][0][1] + company_f['words'][1][1]:.2f}")
            st.metric("Negative Force", f"{company_f['words'][-1][1] + company_f['words'][-2][1]:.2f}")
        
        with col2:
            # SHAP for Company F
            words_f = company_f['words']
            top_pos_f = sorted([w for w in words_f if w[1] > 0], key=lambda x: x[1], reverse=True)[:7]
            top_neg_f = sorted([w for w in words_f if w[1] < 0], key=lambda x: x[1])[:3]
            
            combined_f = top_pos_f + top_neg_f
            combined_f.sort(key=lambda x: x[1])
            
            fig_f = go.Figure(data=[
                go.Bar(
                    y=[w[0] for w in combined_f],
                    x=[w[1] for w in combined_f],
                    orientation='h',
                    marker_color=['#2ecc71' if w[1] > 0 else '#e74c3c' for w in combined_f],
                    text=[f"{w[1]:+.2f}" for w in combined_f],
                    textposition='outside'
                )
            ])
            
            fig_f.update_layout(
                title="Company F: Word Impact",
                xaxis_title="SHAP Value",
                template='plotly_white',
                height=400,
                showlegend=False
            )
            
            st.plotly_chart(fig_f, use_container_width=True)
        
        st.warning(f"""
        **Result:** 
        - VADER: {company_f['vader']:.3f} (very warm)
        - RoBERTa: {company_f['roberta']:.3f} (very warm)
        - **Gap: {abs(company_f['vader'] - company_f['roberta']):.3f}** (both models agree!)
        
        **The strategy worked:** Positive flooding + smart apology choice + joy words = warmest email in dataset.
        """)
        
        st.markdown("---")
        
        # Case Study 2: Company C (Loser)
        st.markdown("### 💔 Case Study #2: Company C - The Disagreement (VADER 0.974, RoBERTa -0.168)")
        
        company_c = shap_results['Company_C']
        
        col1, col2 = st.columns(2)
        
        with col1:
            st.markdown("""
            <div style="background-color: #F3EDFA; padding: 20px; border-radius: 10px;">
            <h4>❌ What Went Wrong</h4>
            <ul>
                <li><b>The killer word:</b> "Unfortunately" with -0.7727 impact (worst in dataset!)</li>
                <li><b>Weak positives:</b> Generic words like "thanks" and "good"</li>
                <li><b>No compensation:</b> Not enough strong positives to balance apology</li>
                <li><b>VADER fooled:</b> Counted positive words, missed rejection context</li>
            </ul>
            </div>
            """, unsafe_allow_html=True)
            
            st.metric("Positive Force", f"+{sum([w[1] for w in company_c['words'] if w[1] > 0]):.2f}")
            st.metric("Negative Force", f"{sum([w[1] for w in company_c['words'] if w[1] < 0]):.2f}")
        
        with col2:
            # SHAP for Company C
            words_c = company_c['words']
            top_pos_c = sorted([w for w in words_c if w[1] > 0], key=lambda x: x[1], reverse=True)[:5]
            top_neg_c = sorted([w for w in words_c if w[1] < 0], key=lambda x: x[1])[:5]
            
            combined_c = top_pos_c + top_neg_c
            combined_c.sort(key=lambda x: x[1])
            
            fig_c = go.Figure(data=[
                go.Bar(
                    y=[w[0] for w in combined_c],
                    x=[w[1] for w in combined_c],
                    orientation='h',
                    marker_color=['#2ecc71' if w[1] > 0 else '#e74c3c' for w in combined_c],
                    text=[f"{w[1]:+.2f}" for w in combined_c],
                    textposition='outside'
                )
            ])
            
            fig_c.update_layout(
                title="Company C: Word Impact",
                xaxis_title="SHAP Value",
                template='plotly_white',
                height=400,
                showlegend=False
            )
            
            st.plotly_chart(fig_c, use_container_width=True)
        
        st.warning(f"""
        **Result:**
        - VADER: {company_c['vader']:.3f} (thinks it's warm!)
        - RoBERTa: {company_c['roberta']:.3f} (knows it's cold!)
        - **Gap: {abs(company_c['vader'] - company_c['roberta']):.3f}** (BIGGEST disagreement in dataset!)
        
        **The lesson:** One terrible word ("unfortunately" -0.77) can destroy an entire email. 
        Transformers learned this is a rejection pattern, while lexicons just count words.
        """)
        
        st.markdown("---")
        
        # Comparison table
        st.markdown("### 📊 Winner vs Loser Comparison")
        
        comparison = pd.DataFrame({
            'Metric': [
                'VADER Score',
                'RoBERTa Score',
                'Model Agreement',
                'Positive Word Sum',
                'Negative Word Sum',
                'Net Impact',
                'Joy Words Used',
                'Worst Word Impact',
                'Best Word Impact'
            ],
            'Company F (Winner)': [
                f"{company_f['vader']:.3f}",
                f"{company_f['roberta']:.3f}",
                f"{abs(company_f['vader'] - company_f['roberta']):.3f} ✅",
                "+2.67 🏆",
                "-0.45 ✅",
                "+2.22 🏆",
                "2 (wish, best)",
                "-0.17 (disappointing)",
                "+0.42 (thank)"
            ],
            'Company C (Loser)': [
                f"{company_c['vader']:.3f}",
                f"{company_c['roberta']:.3f}",
                f"{abs(company_c['vader'] - company_c['roberta']):.3f} ❌",
                "+1.03",
                "-1.59 ❌",
                "-0.56 ❌",
                "1 (good)",
                "-0.77 (unfortunately) ☠️",
                "+0.18 (thanks)"
            ]
        })
        
        st.dataframe(comparison, use_container_width=True, hide_index=True)
        
        st.markdown("""
        **The Contrast:**
        - **Positive sum:** Winner has 2.6× more positive force
        - **Negative sum:** Winner has 72% less negative force
        - **Net impact:** Winner is +2.78 points better
        - **Model agreement:** Winner has models aligned, loser has 1.14 gap
        
        **Bottom line:** Company F's template is objectively, measurably better.
        """)

else:  # Try It Yourself
    st.title("🧪 Try It Yourself")
    st.markdown("### Analyze your own rejection email or create a better template")
    
    tab1, tab2 = st.tabs(["🔍 Analyze", "✍️ Rewrite"])
    
    with tab1:
        st.markdown("## 📧 Email Analyzer")
        st.markdown("Paste any rejection email to see how it scores:")
        
        # Example selector
        example = st.selectbox(
            "Or try an example:",
            ["", "❌ Cold Email", "😐 Generic Email", "✅ Warm Email"]
        )
        
        example_texts = {
            "❌ Cold Email": """Dear [Name],

Thanks for applying to [Role].

Unfortunately, we have decided not to move forward with your application at this time. We're sorry we couldn't offer you a position.

Thanks again,
[Team]""",
            "😐 Generic Email": """Hi [Name],

Thank you for your interest in the [Role] position.

After careful consideration, we have decided to move forward with other candidates whose qualifications more closely match our needs.

We wish you success in your job search.

Best regards,
[Recruiter]""",
            "✅ Warm Email": """Dear [Name],

Thank you for your interest in [Role] and for taking the time to interview with us.

We were impressed by your background in [field] and appreciated learning about your experience with [topic]. While we're moving forward with another candidate whose skills more closely align with this specific role, we want you to know this was a very competitive process.

We encourage you to watch for future opportunities that match your expertise.

We wish you the best in your search!

Warm regards,
[Name]"""
        }
        
        text_input = st.text_area(
            "Paste email here:",
            value=example_texts.get(example, ""),
            height=250,
            placeholder="Dear [Name],\n\nThank you for applying..."
        )
        
        if st.button("🚀 Analyze Email", type="primary"):
            if len(text_input.strip()) < 10:
                st.warning("Please paste an email first!")
            else:
                results = analyze_text(text_input)
                
                # Results
                st.markdown("---")
                st.markdown("### 📊 Results")
                
                # Gauge and metrics
                col1, col2 = st.columns([1, 1])
                
                with col1:
                    # Simple gauge visualization
                    score = results['score']
                    if score >= 0.95:
                        st.success(f"### ✅ Very Warm\n**Score: {score:.3f}**")
                    elif score >= 0.85:
                        st.info(f"### 😊 Warm\n**Score: {score:.3f}**")
                    elif score >= 0.60:
                        st.warning(f"### 😐 Neutral\n**Score: {score:.3f}**")
                    else:
                        st.error(f"### ❌ Cold\n**Score: {score:.3f}**")
                
                with col2:
                    st.markdown("#### 📈 Breakdown")
                    st.metric("Joy Words", results['joy'])
                    st.metric("Positive Words", results['positive'])
                    st.metric("Apologies", results['apology'])
                
                # Analysis
                st.markdown("### 🔍 Analysis")
                
                problems = []
                if results['apology'] > 1:
                    problems.append(f"⚠️ Too many apologies ({results['apology']}). Limit to 0-1.")
                elif results['apology'] == 1 and results['positive'] < 4:
                    problems.append(f"⚠️ Not enough positives for apology. Need 4+, have {results['positive']}.")
                
                if results['joy'] == 0:
                    problems.append("⚠️ No joy words. Add 'wish', 'hope', or 'best'.")
                
                if results['positive'] < 4:
                    problems.append(f"⚠️ Only {results['positive']} positive words. Aim for 6+.")
                
                if problems:
                    st.error("**Problems Detected:**")
                    for p in problems:
                        st.markdown(f"- {p}")
                else:
                    st.success("✅ No major problems detected!")
                
                # Recommendations
                st.markdown("### 💡 Recommendations")
                if results['score'] >= 0.95:
                    st.success("This is an excellent rejection email! It balances warmth with professionalism.")
                elif results['score'] >= 0.85:
                    st.info("This is a warm email. Consider adding more joy words to make it exceptional.")
                else:
                    st.warning("""
                    This email could be warmer. Try:
                    - Adding joy words like "wish you the best"
                    - Increasing positive words to 6+
                    - Following the 4:1 rule (4 positives per apology)
                    - Avoiding "unfortunately" and "regret"
                    """)
    
    with tab2:
        st.markdown("## ✍️ Email Rewriter")
        st.markdown("Paste a cold email template and we'll suggest improvements:")
        
        rewrite_input = st.text_area(
            "Paste template here:",
            height=200,
            placeholder="Dear [Name],\n\nThanks for applying..."
        )
        
        if st.button("🔍 Analyze & Improve", type="primary"):
            if len(rewrite_input.strip()) < 10:
                st.warning("Please paste an email first!")
            else:
                original = analyze_text(rewrite_input)
                
                # Simple improvements
                improved = rewrite_input
                suggestions = []
                
                # Remove "unfortunately"
                if 'unfortunately' in improved.lower():
                    improved = re.sub(r'\b[Uu]nfortunately,?\s*', '', improved)
                    suggestions.append("✅ Removed 'unfortunately' (saves up to -0.77 negative impact!)")
                
                # Replace "sorry" with "disappointed"
                if 'sorry' in improved.lower():
                    improved = re.sub(r"\bsorry\b", "disappointed", improved, flags=re.IGNORECASE)
                    suggestions.append("✅ Changed 'sorry' to 'disappointed' (+64% less negative)")
                
                # Replace "regret" with "disappointing"
                if 'regret' in improved.lower():
                    improved = re.sub(r"\bregret\b", "find it disappointing", improved, flags=re.IGNORECASE)
                    suggestions.append("✅ Changed 'regret' to 'disappointing' (+77% less negative)")
                
                # Add joy words
                if original['joy'] == 0 and 'wish you' not in improved.lower():
                    improved += "\n\nWe wish you the best in your job search!"
                    suggestions.append("✅ Added joy words to closing (+0.23 impact)")
                
                # Add specific positive
                if original['positive'] < 4 and 'impressed' not in improved.lower() and 'appreciate' not in improved.lower():
                    lines = improved.split('\n')
                    for i, line in enumerate(lines):
                        if any(word in line.lower() for word in ['dear', 'hi ', 'hello']):
                            lines.insert(i+1, "\nWe appreciate the time you invested in your application and were impressed by your background.\n")
                            break
                    improved = '\n'.join(lines)
                    suggestions.append("✅ Added 'appreciate' and 'impressed' (+1.05 combined impact)")
                
                # Enhance "thank" to "appreciate"
                if improved.count('thank') > 0 and 'appreciate' not in improved.lower():
                    improved = improved.replace('Thank you', 'We appreciate your', 1)
                    suggestions.append("✅ Enhanced 'thank' to 'appreciate' (+300% more impact)")
                
                improved_analysis = analyze_text(improved)
                
                # Show results
                st.markdown("---")
                st.markdown("### 📊 Results")
                
                col1, col2, col3 = st.columns(3)
                with col1:
                    st.metric("Before", f"{original['score']:.3f}")
                with col2:
                    st.metric("After", f"{improved_analysis['score']:.3f}")
                with col3:
                    change = improved_analysis['score'] - original['score']
                    st.metric("Improvement", f"{change:+.3f}", delta=f"{change/abs(original['score'])*100:+.0f}%")
                
                st.markdown("### ✨ Changes Applied")
                if suggestions:
                    for s in suggestions:
                        st.markdown(f"- {s}")
                else:
                    st.success("Email is already optimal!")
                
                st.markdown("### 📧 Comparison")
                col1, col2 = st.columns(2)
                
                with col1:
                    st.markdown("**Original:**")
                    st.text_area("", value=rewrite_input, height=300, disabled=True, label_visibility="collapsed")
                
                with col2:
                    st.markdown("**Improved:**")
                    st.text_area("", value=improved, height=300, disabled=True, label_visibility="collapsed")
                
                # Show the math
                if suggestions:
                    st.info(f"""
                    **💡 Why This Works:**
                    
                    Based on SHAP analysis of 14 real emails:
                    - Removing "unfortunately" = up to +0.77 improvement
                    - Swapping "sorry" for "disappointed" = +0.30 improvement
                    - Adding "appreciate" instead of "thank" = +0.40 improvement
                    - Adding joy words = +0.20+ improvement
                    
                    **Total potential improvement: {change:+.2f} points!**
                    """)

# Run
if __name__ == "__main__":
    pass