Product Management Portfolio Project | Data-Driven Customer Experience Analysis
Analyzed 7,800 unique customer reviews (2,600 per platform) across DoorDash, UberEats, and GrubHub to extract actionable insights on customer pain points and satisfaction drivers. Built an AI-powered pipeline that automatically identifies and categorizes customer feedback into 18 distinct aspects with sentiment classification.
Key Findings:
- 70.7% negative sentiment across platforms indicates critical service gaps
- Customer support is the #1 complaint category (1,867 negative mentions)
- Delivery reliability is the #2 pain point (1,246 negative mentions)
- GrubHub shows highest negative sentiment (64.2%) vs DoorDash (54.3%) and UberEats (60.7%)
This project aims to understand the pain points of customers who utilize these Food delivery platforms by:
- Automatically categorizing vague feedback into specific actionable categories
- Quantifying sentiment across 18 distinct customer experience dimensions
- Identifying platform-specific weaknesses to prioritize product improvements
- Scaling qualitative insights from thousands of reviews in minutes, not months
- Each customer review is first passed through a RoBERTa model custom fine-tuned by me with 94% accuracy for aspect-based sentiment analysis. RoBERTa Fine-Tuned Model published at Hugging Face
- The model automatically identifies all relevant parent aspects present in the text and determines the sentiment (positive, neutral, negative) for each detected aspect.
- This stage provides a structured overview of which major experience areas (e.g., food, delivery, service, price, interface, overall) are discussed and how customers feel about them.
- Example:
- Input: "The pizza was cold and soggy, but the app was easy to use."
- Output:
{ 'food': 'negative', 'interface': 'positive' }
- For each detected aspect in a review, a large language model (Qwen 2.5 14B) further breaks down the feedback into fine-grained subcategories for deeper insight.
- This step helps clarify exactly what part of an aspect (e.g., food quality vs. food taste) is being praised or criticized.
- Subcategories are extracted using strict definitions and boundary rules to ensure non-overlapping, actionable categories.
- Example:
- For "food" aspect in "The pizza was cold and soggy", returns:
[food_quality, food_freshness] - For "interface" aspect in "The app was easy to use", returns:
[app_usability]
- For "food" aspect in "The pizza was cold and soggy", returns:
The following parent aspects and their subcategories are used (with strict definitions):
- food_quality: Physical condition and preparation of food ONLY. Temperature, texture, cooking level, consistency.
- food_taste: Flavor and seasoning ONLY. Taste, spice, saltiness, sweetness, blandness.
- food_freshness: Age and freshness of ingredients ONLY. Spoilage, expired items, wilted produce.
- food_presentation: Visual appearance and plating ONLY. Looks, arrangement, garnish.
- delivery_speed: Time taken for delivery ONLY. Fast, slow, late, early, on-time.
- delivery_reliability: Accuracy and dependability ONLY. Wrong/missing items, address issues.
- driver_behavior: Driver’s conduct and professionalism ONLY. Politeness, rudeness, communication.
- packaging_quality: Physical packaging condition ONLY. Sealed, damaged, leaking, secure containers.
- customer_support: Help from support team ONLY. Response to complaints, refunds, issue resolution.
- staff_attitude: Restaurant staff behavior ONLY (not driver). Kitchen staff, order takers, management.
- responsiveness: Communication speed ONLY. Promptness from restaurant or support.
- value_for_money: Perceived quality vs. cost ONLY. Portion size, worth it.
- fees_charges: Delivery/service fees ONLY. Extra charges, transparency.
- discounts_promotions: Deals, coupons, and their application.
- pricing_fairness: Billing accuracy ONLY. Overcharge, wrong price.
- app_usability: Ease of use, bugs ONLY. App crashes, navigation, clarity.
- navigation: Finding items or restaurants ONLY. Search, menu organization.
- app_features: Specific app features ONLY. Tracking, payment, customization.
- overall_satisfaction: Vague or general feedback not specific to any above aspect.
Each subcategory is defined with clear boundaries to ensure high-quality, non-overlapping labeling. For more detail on subcategory definitions and keywords, see the aspect_extraction.py file.
Customer Review
↓
┌──────────────────────────────┐
│ FABSA RoBERTa (GPU) │ Detect aspects + sentiment
│ "What is discussed & how?" │ → {food: negative, interface: positive}
└──────────────────────────────┘
↓
┌──────────────────────────────┐
│ Qwen 2.5 14B (GPU) │ Extract subcategories for each aspect
│ "What about this aspect?" │ → food: [food_quality, food_freshness]
└──────────────────────────────┘
↓
Highly Actionable Insights
| Platform | Negative % | Positive % | Top Pain Point |
|---|---|---|---|
| GrubHub | 74.8% | 26.0% | Customer support (811 mentions) |
| UberEats | 67.2% | 24.8% | Customer support (659 mentions) |
| DoorDash | 65.4% | 32.9% | Customer support (589 mentions) |
-
Customer Support (2,059 mentions, 85% negative)
- Recommendation: Implement live chat + faster refund automation
-
App Usability (1,429 mentions, 78% negative)
- Recommendation: Address crashes, simplify checkout flow
-
Delivery Reliability (1,409 mentions, 82% negative)
- Recommendation: Improve order accuracy tracking, reduce wrong/missing items
-
Fees & Charges (1,132 mentions, 88% negative)
- Recommendation: Increase pricing transparency, bundle fees
-
Value for Money (992 mentions, 74% negative)
- Recommendation: Right-size portions, adjust pricing tiers
- Delivery Speed shows highest positive sentiment (42% positive)
- Discounts/Promotions drive customer satisfaction when present
- Food Quality receives praise when executed well (fresh, hot, properly cooked)
Food (4)
food_quality: Temperature, texture, cooking levelfood_taste: Flavor, seasoning, spicefood_freshness: Ingredient age, spoilagefood_presentation: Visual appearance, plating
Delivery (4)
delivery_speed: Delivery timedelivery_reliability: Order accuracydriver_behavior: Professionalism, communicationpackaging_quality: Container condition, leaks
Service (3)
customer_support: Issue resolution, refundsstaff_attitude: Restaurant staff behaviorresponsiveness: Communication speed
Price (4)
value_for_money: Quality vs. costfees_charges: Delivery/service feesdiscounts_promotions: Deals, couponspricing_fairness: Billing accuracy
Interface (3)
app_usability: Ease of use, bugsnavigation: Finding items/restaurantsapp_features: Tracking, payment, customization
Overall (1)
overall_satisfaction: General vague feedback
- Python 3.12+
- NVIDIA GPU (16GB+ VRAM recommended)
- Ollama installed
# Clone repository
git clone <your-repo-url>
cd <repo-name>
# Install dependencies
pip install pandas transformers torch requests
# Install Ollama and download model
curl -fsSL https://ollama.com/install.sh | sh
ollama serve &
ollama pull qwen2.5:14b-instructpython run_analysis.pyOutput:
- Individual platform results:
output/doordash_analysis.csv - Combined results:
output/complete_analysis_YYYYMMDD_HHMMSS.csv - Checkpoints saved every 500 reviews for fault tolerance
🎯 Multi-Platform Aggregator Dashboard
Platform-Specific Dashboards:
- Real-time KPI Tracking: Average rating, top pain points, correlation analysis
- Interactive Visualizations:
- Parent aspect distribution bar charts
- Sentiment breakdown donut charts
- Subcategory treemaps with intensity heatmaps
- Co-occurrence correlation matrices
- NEW: Platform comparison pyramid (side-by-side analysis)
- Dynamic Filtering: Date range, aspect filtering, and platform selectors
- Platform Comparison: Visual benchmarking with brand colors (DoorDash red, UberEats green, GrubHub orange)
# Navigate to any dashboard directory
cd dashboard-aggregator/
# Start local server
python -m http.server 8000
# Open in browser
# http://localhost:8000├── data/ # Input review datasets
│ ├── doordash_customer_reviews.csv
│ ├── ubereats_customer_reviews.csv
│ └── grubhub_customer_reviews.csv
├── src/ # Source code
│ ├── aspect_extraction.py # LLM-based subcategory extraction
│ ├── sentiment_analyzer.py # RoBERTa sentiment classification
│ └── pipeline.py # End-to-end pipeline orchestration
├── output/ # Analysis results
│ └── complete_analysis_*.csv # Final output with all insights
├── dashboard-aggregator/ # Multi-platform analytics dashboard
├── doordash-dashboard/ # DoorDash-specific dashboard
├── dashboard-ubereats/ # UberEats-specific dashboard
├── dashboard-grubhub/ # GrubHub-specific dashboard
├── run_analysis.py # Main execution script
└── README.md # This file
# Input Review
"The pizza arrived cold but the driver was very polite"
# Pipeline Output
{
'food_quality': {
'sentiment': 'negative',
'confidence': 0.94,
'parent_aspect': 'food'
},
'driver_behavior': {
'sentiment': 'positive',
'confidence': 0.89,
'parent_aspect': 'delivery'
}
}Output CSV Format:
| review_id | subcategory | sentiment | confidence | platform |
|---|---|---|---|---|
| abc123 | food_quality | negative | 0.94 | doordash |
| abc123 | driver_behavior | positive | 0.89 | doordash |
- Data-driven decision making: Quantified customer pain points from unstructured feedback
- Cross-functional thinking: Bridged business needs with ML implementation
- Prioritization: Ranked features by customer impact (support > app > delivery)
- Competitive analysis: Benchmarked 3 platforms to identify gaps
- Machine Learning: Multi-label classification, sentiment analysis, LLM prompting
- Python Development: Pipeline architecture, error handling, checkpointing
- GPU Optimization: CUDA acceleration, batch processing
- Data Analysis: Pandas, statistical summarization, insight extraction
- Customer Experience: Understanding CX metrics and pain point categorization
- Food Delivery: Platform-specific workflows and user journeys
- NLP: Aspect-based sentiment analysis, semantic classification
- 7,800 reviews processed (2,600 per platform)
- 13,868 aspect-sentiment pairs extracted
- ~0.5s per review on NVIDIA A40 GPU
- 18 distinct subcategories identified
- 3 sentiment classes (positive, neutral, negative)
Sentiment Breakdown:
- Negative: 9,801 (70.7%)
- Positive: 3,917 (28.2%)
- Neutral: 150 (1.1%)
- Time-series analysis: Track sentiment trends over months
- Root cause clustering: Group similar complaints for deeper insights
- Predictive modeling: Forecast churn based on sentiment patterns
- Real-time dashboard: Live monitoring of customer feedback
- Recommendation engine: Suggest specific product fixes per platform
MIT License - See LICENSE file for details
Product Manager | Data Analyst
This project demonstrates my ability to:
- Translate business problems into technical solutions
- Extract actionable insights from large-scale unstructured data
- Build production-grade ML pipelines
- Communicate findings to stakeholders
anudeepreddynarala2@gmail.com https://www.linkedin.com/in/anudeep-narala/
Built with Qwen 2.5 14B, FABSA RoBERTa, PyTorch, and passion for customer experience.