• 🔤 NLP Pipeline: tokenization, lemmatization, POS tagging, stopword filtering
• 🔍 Topic Modeling with NMF – reveals semantic clusters
• 🧮 Multiple ML Classifiers
  • Logistic Regression (≈95% accuracy)
  • Decision Tree
  • Random Forest
  • Gaussian Naive Bayes
    (LR was the top performer.)
• 🎨 Rich EDA: histograms, distributions, n-grams, word clouds
• ⚙️ Complete Inference Pipeline: ready for deployment
• 🧪 Interpretability: TF-IDF feature importance, topic keywords
• 📚 Full research documentation in PDF

Architecture: Ingestion → Processing → Topic Modeling → ML Classifier → Prediction

Model Workflow: Data → Clean → TF-IDF → NMF → Classifier → Results

Word Cloud of Most Frequent Complaint Tokens

From the dataset EDA:

• Strong presence of tokens like payment, credit, account, dispute, reporting
• Complaints map naturally into financial service categories
• Topic-token alignment validated using NMF factors
  :contentReference[oaicite:3]{index=3}

git clone https://github.com/kalpthakkar/JobPilot-AI.git
cd jobpilot_ai

python -m venv .venv_resolveai

# macOS/Linux
source .venv_resolveai/bin/activate


# Windows
.venv_resolveai\Scripts\activate
# Powershell
.venv_resolveai\Scripts\Activate.ps1

jupyter lab src/notebook.ipynb

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.decomposition import NMF
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

tfidf = TfidfVectorizer(max_df=0.95, min_df=2, ngram_range=(1,2))
X = tfidf.fit_transform(df['clean_complaint'])

nmf = NMF(n_components=5, random_state=42)
W = nmf.fit_transform(X)

y = W.argmax(axis=1)

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

clf = LogisticRegression(max_iter=1000)
clf.fit(X_train, y_train)

print(classification_report(y_test, clf.predict(X_test)))

Overall Model Performance

• ✔ LR chosen as final classifier
• ✔ Well-separated NMF topics improved class clarity
• ✔ Topic-word alignment validated

• Fine-tuned transformer models (GPT, BERT, DistilBERT)
• Real-time ingestion with Kafka
• FastAPI inference service
• Active learning loop from agent feedback
• Multi-modal complaint classification (voice → transcripts → NLP)

Q: What dataset did you use?
A dataset of 78,313 customer complaints, with 22 columns, provided in JSON format.

Q: Why NMF for topic modeling?
NMF provides sparse, interpretable topics that align well with complaint categories.

Q: Why did Logistic Regression perform best?
High-dimensional TF-IDF vectors naturally favor linear decision boundaries.

Q: Can this model run in production?
Yes, TF-IDF + LR is fast, light, and easily containerizable.

Q: How is bias or PII handled?
Remove PII (names, IDs, emails) and evaluate demographic fairness before launch.

This Machine Learning work is developed by Kalp Thakkar. References and full experiments in report.pdf and slides

Bird, Klein & Loper - Natural Language Processing with Python (NLTK).
Pedregosa et al. - Scikit-learn.
Lee & Seung - NMF (Nature, 1999).

See full bibliography in this report

• Fork repository

  https://github.com/kalpthakkar/ReSolveAI-AI-Complaint-Classification-Engine/fork

• Create a branch:

  git checkout -b feature-xyz

• Commit your changes

  # Stage changes
  git add .
  
  # Commit
  git commit -m "Your message"
  
  # Push the new branch
  git push -u origin feature-xyz

• Submit a Pull Request

  gh pr create --fill

ReSolveAI showcases how classical NLP + topic modeling + supervised ML can deliver real, measurable impact in customer complaint handling pipelines. This repository demonstrates research quality, engineering quality, and practical industry relevance - all in one project.

For any inquiries or support, please contact:

• Kalp Thakkar - kalpthakkar2001@gmail.com
• GitHub: kalpthakkar
• LinkedIn: kalpthakkar