🧠 Employee Sentiment Analysis — Work Report

Springer Capital | Data Analysis & LLM Intern | Aug 2025 – Nov 2025 | Remote

> Fine-tuned DistilBERT pipeline for employee sentiment scoring, 30-day flight-risk modeling (R²=0.81), and HR intelligence across 10,000+ records. Fully containerized in Docker.

---

📌 Project Overview

End-to-end NLP pipeline analyzing 10,000+ employee email records spanning 24 months (2010–2011). The system classifies sentiment per employee per month, identifies flight-risk signals using a 30-day rolling model, and delivers scored outputs to the HR/product team.

Business outcome: Seaborn EDA dashboards were packaged into a management deck and directly drove two concrete HR interventions based on identified flight-risk patterns.

---

🏗️ Pipeline Architecture

Raw Email Data
    │
    ▼

Data Cleaning          ← remove nulls, duplicates, encoding noise
    │
    ▼
Preprocessing          ← lowercase, punctuation removal,
stopword filtering, tokenization
    │
    ▼
Feature Engineering    ← text_length | word_count | month_year
    │
    ▼
DistilBERT Fine-tuning ← ~97% of BERT-base accuracy
40% smaller | 60% faster inference
    │
    ▼
Sentiment Scoring      ← per-employee monthly aggregation
    │
    ▼
Flight-Risk Model      ← 30-day rolling window
rule-based heuristics + regression ensemble
    │
    ▼
Output CSVs            ← scored outputs → product team

---

📊 EDA & Feature Engineering

• Dataset: 10,000+ employee email records (2010–2011, 24 months)
• Visualization: Seaborn distribution plots, time-series sentiment trends, employee heatmaps
• Business impact: Dashboards packaged into a management deck → drove 2 HR interventions

Engineered Features:

Feature	Description
text_length	Character count per message
word_count	Whitespace-tokenized word count
month_year	Temporal aggregation key for trend analysis

---

🤖 Model: DistilBERT Fine-Tuning

Model Comparison

Model	Relative Accuracy	Speed	Size
BERT-base	100% (baseline)	Slow	Full
DistilBERT (chosen)	~97% of BERT	60% faster	40% smaller
TextBlob	~60-70% on domain text	Fast	Tiny

Why DistilBERT: Context-aware transformer handles workplace-specific language, negations ("not a good experience"), and domain terminology accurately. TextBlob — trained on movie reviews — systematically fails on business email tone.

Why not full BERT: High latency and compute cost; the marginal accuracy gain was not justifiable at production scale.

---

🚨 Flight-Risk Model

Architecture

30-day rolling window combining two components:

1. Rule-based heuristics — flag employees with ≥ 4 negative sentiment signals in any 30-day window
2. Regression ensemble — trained on engineered features to predict future negative signal count

Validation Metrics

Metric	Value
R² Score	0.81
MSE	1.34

Scored outputs were handed off to the product team for downstream HR action.

Flagged Employees (Top 10 by Negative Signal Count)

Employee	Window Start	Window End	Negative Count
rhonda.denton@enron.com	2010-01-24	2010-02-23	5
bobette.riner@ipgdirect.com	2010-01-09	2010-02-08	4
don.baughman@enron.com	2009-12-24	2010-01-23	4
eric.bass@enron.com	2009-12-11	2010-01-10	4
john.arnold@enron.com	2009-12-28	2010-01-27	4
johnny.palmer@enron.com	2009-12-20	2010-01-19	4
kayne.coulter@enron.com	2009-12-19	2010-01-18	4
lydia.delgado@enron.com	2009-12-14	2010-01-13	4
patti.thompson@enron.com	2009-12-16	2010-01-15	4
sally.beck@enron.com	2009-12-17	2010-01-16	4

---

🏆 24-Month Sentiment Summary

Top Positive Employees

Rank	Employee	Appearances in Top 3
1	Eric Bass	10
2	Bobette Riner	8
3	Johnny Palmer	7

Top Negative Employees

Rank	Employee	Appearances in Bottom 3
1	John Arnold	10
2	Don Baughman	9
3	Kayne Coulter	7

---

🐳 Running the Project (Docker)

# Build
docker build -t employee-sentiment .

# Run full pipeline
docker run employee-sentiment

# Run with local output mount
docker run -v $(pwd)/outputs:/app/outputs employee-sentiment

No local Python environment needed. All dependencies locked in requirements.txt.

---

🛠️ Tech Stack

Layer	Technology
NLP Model	DistilBERT (HuggingFace Transformers)
EDA & Visualization	Pandas, Seaborn, Matplotlib
Feature Engineering	Scikit-learn, Custom text features
Flight-Risk Model	Scikit-learn regression ensemble
Containerization	Docker
Notebook	Jupyter

---

📚 LLM Training Track

As part of the internship curriculum, a parallel LLM study track was completed alongside this data project:

LangChain

• Prompt chaining and template management
• Retrieval-Augmented Generation (RAG) workflows
• Tool calling and agent construction patterns

LangGraph

• Stateful multi-agent graph orchestration
• Node and edge design for complex agent flows
• State management across multi-step reasoning

These concepts directly informed subsequent agentic projects (Agentic RAG Telegram Bot with LangGraph, OpenSearch, Redis, Airflow).

---

📁 Repository Structure

Employee-Sentimental-analysis/
  ├── employee_sentiment_analysis.ipynb   ← Full pipeline notebook
  ├── labeled_messages_with_scores.csv    ← Scored output (product team)
  ├── test.csv                            ← Test split
  ├── requirements.txt                    ← Locked dependencies
  ├── Dockerfile                          ← Reproducible container
  ├── .dockerignore
  ├── Final_Report.docx                   ← Management deck
  ├── visualizations/                     ← Seaborn EDA dashboards
  └── README.md

---

🏢 About This Project

Completed as part of the Data Analysis & LLM Internship at Springer Capital Aug to Nov 2025 · Remote

---

By Shaswat Raj

---