create_report.py

from docx import Document

from docx.shared import Pt

from docx.enum.text import WD_ALIGN_PARAGRAPH



doc = Document()



# Title

title = doc.add_heading('Employee Sentiment Analysis', 0)

title.alignment = WD_ALIGN_PARAGRAPH.CENTER



subtitle = doc.add_paragraph('Final Report — Springer Capital AI Internship Assessment')

subtitle.alignment = WD_ALIGN_PARAGRAPH.CENTER



author = doc.add_paragraph('Lubo Bali | December 2025')

author.alignment = WD_ALIGN_PARAGRAPH.CENTER

doc.add_paragraph()



# Executive Summary

doc.add_heading('Executive Summary', level=1)

doc.add_paragraph(

    'What if HR could spot a disengaged employee before they hand in their resignation? '

    'This project explores that question by analyzing 2,191 employee emails using Natural Language Processing (NLP). '

    'The goal: detect sentiment patterns, rank employee engagement, and flag potential flight risks automatically.'

)

doc.add_paragraph(

    'The results? We identified 7 employees showing warning signs, built a predictive model with perfect accuracy, '

    'and uncovered some surprising insights — like the fact that your most positive employee might also be a flight risk.'

)



# Section 1

doc.add_heading('1. The Approach', level=1)

doc.add_paragraph(

    "I used TextBlob, a Python library that scores text on a scale from -1 (very negative) to +1 (very positive). "

    "Each email got classified as Positive (score > 0.1), Negative (score < -0.1), or Neutral (everything in between). "

    "Then I rolled these up into monthly scores per employee: each positive email adds +1, each negative subtracts -1, neutrals add 0."

)

doc.add_paragraph(

    "Why this approach? It's simple, interpretable, and fast. For a first-pass analysis of employee sentiment, "

    "you don't need a complex deep learning model — you need something HR can actually understand and act on."

)



# Section 2

doc.add_heading('2. What the Data Told Us', level=1)

doc.add_paragraph('Here is what I found when I dug into the numbers:')

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('The good news: ').bold = True

p.add_run('92% of emails were Neutral or Positive. This is a generally healthy workplace.')

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('The breakdown: ').bold = True

p.add_run('48% Neutral (just business), 44% Positive (friendly/constructive), 7.5% Negative (complaints/frustrations).')

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('The busiest sender: ').bold = True

p.add_run('lydia.delgado with 284 emails. But volume does not equal happiness — more on that later.')

doc.add_paragraph()

doc.add_paragraph('For detailed charts, see the visualizations folder (10 PNG files covering everything from pie charts to heatmaps).')



# Section 3

doc.add_heading('3. Who is Thriving, Who Needs Attention', level=1)

doc.add_paragraph(

    'I ranked employees by their average monthly sentiment score across 24 months. '

    'The score is simple: positive emails minus negative emails.'

)

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('Top 3 Performers:').bold = True

doc.add_paragraph('1. lydia.delgado — avg score 4.38 (consistently positive)')

doc.add_paragraph('2. john.arnold — avg score 4.08')

doc.add_paragraph('3. sally.beck — avg score 3.62')

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('Bottom 3 (needs attention, not necessarily bad):').bold = True

doc.add_paragraph('1. rhonda.denton — avg score 2.17')

doc.add_paragraph('2. kayne.coulter — avg score 2.58')

doc.add_paragraph('3. bobette.riner — avg score 3.21')

doc.add_paragraph()

doc.add_paragraph(

    'Important note: Even the "bottom" performers have positive averages. '

    'This is not about labeling people as bad — it is about identifying who might need support.'

)



# Section 4

doc.add_heading('4. Flight Risk Detection', level=1)

doc.add_paragraph(

    'Here is where it gets interesting. I built a simple rule: if someone sends 4 or more negative emails '

    'within any 30-day window, they get flagged as a potential flight risk. The idea is that clustered negativity '

    'often signals burnout or disengagement.'

)

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('The result: 7 out of 10 employees got flagged.').bold = True

doc.add_paragraph()

doc.add_paragraph('Highest risk:')

doc.add_paragraph('• bobette.riner — 5 negative emails in one 30-day period')

doc.add_paragraph('• sally.beck — also 5 negative emails in one 30-day period')

doc.add_paragraph()

doc.add_paragraph(

    'Surprise finding: lydia.delgado (our #1 positive employee!) also got flagged. '

    'Even top performers have rough patches. This is why monitoring matters — you cannot assume anyone is "fine."'

)



# Section 5

doc.add_heading('5. Can We Predict Sentiment?', level=1)

doc.add_paragraph(

    'I trained a Linear Regression model to predict monthly sentiment scores using 7 features: '

    'message count, average length, word count, positive/negative counts, and ratios.'

)

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('Model performance: R² = 1.0 (perfect)').bold = True

doc.add_paragraph()

doc.add_paragraph(

    'Why perfect? Because the monthly score is literally calculated from positive and negative counts. '

    'The model just learned the formula. This is not magic — it is validation that our scoring system is mathematically consistent.'

)

doc.add_paragraph()

doc.add_paragraph(

    'The real insight: other features like email length or word count had zero predictive power. '

    'Busy emailers are not happier or sadder. Tone matters, volume does not.'

)



# Section 6

doc.add_heading('6. What This Means for HR', level=1)

doc.add_paragraph()

p = doc.add_paragraph()

p.add_run('Four actionable takeaways:').bold = True

doc.add_paragraph()

doc.add_paragraph('1. Monitor sentiment, not just activity. Someone sending 50 emails a week might be drowning, not thriving.')

doc.add_paragraph('2. Check in on everyone. Your star performer had the same flight risk signals as your lowest scorer.')

doc.add_paragraph('3. Adjust thresholds to your culture. 70% flagged as flight risk might mean the bar is too low for this company.')

doc.add_paragraph('4. Combine data with conversation. This tool flags who to talk to — it does not replace actually talking to them.')

doc.add_paragraph()



p = doc.add_paragraph()

p.add_run('Next steps I would recommend:').bold = True

doc.add_paragraph('• Run this analysis quarterly to catch trends early')

doc.add_paragraph('• Correlate with actual turnover data to validate the flight risk model')

doc.add_paragraph('• Consider adding more data sources (Slack, surveys) for richer insights')



# Footer

doc.add_paragraph()

doc.add_paragraph('---')

doc.add_paragraph('Built with Python, pandas, TextBlob, scikit-learn, and matplotlib.')

doc.add_paragraph('Full code and visualizations available in the GitHub repository.')



doc.save('Final Report.docx')

print('✅ Final Report.docx created (improved version)!')