🚀 RevSense – Customer Feedback Analytics & Sentiment Insights Platform

Revsense is an end-to-end AI-powered platform that extracts, processes, categorizes, and visualizes customer reviews from multi-source platforms to generate actionable business insights. It combines advanced NLP, real-time automation, and LLM-based reasoning to transform raw user feedback into strategic decision-making tools.

---

📌 Key Features

• 🔁 Automated daily scraping of reviews from platforms like Twitter, Play Store, App Store, and MouthShut using a scheduled Cron Job.
• 🧹 NLP-based text preprocessing with spaCy, including tokenization, punctuation removal, and stopword filtering.
• 🧠 Semantic embedding generation using Sentence-BERT (nlptown/bert-base-multilingual-uncased-sentiment).
• 🧭 5-level hierarchical categorization with 169+ unique leaf-level classes using cosine similarity.
• 📊 Sentiment analysis using a fine-tuned BERT model (Very Negative → Very Positive).
• 📝 Category-wise summarization for thematic understanding.
• 🔁 LLM chunking strategy for passing optimized summaries to Gemini LLM.
• ⚡ High-speed data retrieval with Redis database for efficient date-range queries.
• ⚙️ FastAPI backend for API access and front-end integration.
• 📈 Interactive visualization dashboard for stakeholders.

---

✅ Methodology & Workflow: Point-wise Explanation

1. 🕒 Daily Review Scraping

   • A cron job runs every night to scrape reviews automatically from platforms like Twitter, Play Store, App Store, and MouthShut.

2. 🧹 Review Preprocessing

   • Reviews are cleaned using:
     • Tokenization
     • Stopword removal (using spaCy)
     • Lowercasing
     • Removing punctuation

3. 🧠 Embedding Generation

   • Preprocessed text is passed to Sentence-BERT (nlptown/bert-base-multilingual-uncased-sentiment)
   • This generates dense semantic vectors (embeddings) for each review

4. 🧭 Hierarchical Category Classification

   • We use cosine similarity between review embedding and category embeddings
   • Reviews are classified up to 5 levels of category hierarchy
   • Total of 169 unique leaf categories (e.g., App crash, Delivery delay, Taste quality)

5. 🔢 Category Count & Tracking

   • We count reviews per category, helping to identify which issues occur most frequently

6. 📊 Sentiment Analysis

   • Using BERT model to label reviews as:
     • Very Negative
     • Negative
     • Neutral
     • Positive
     • Very Positive

7. 📝 Summary Generation

   • For each category, a summary of the reviews is created
   • Helps in quickly understanding the nature of feedback

8. 🧩 Chunking for LLM Processing

   • The summaries are divided into chunks before passing to Gemini LLM
   • This helps manage token size limits and speeds up processing time

9. 📈 Insight Aggregation & Visualization

   • Insights are grouped by category and sentiment
   • Displayed on a dashboard for stakeholders to interpret and act on

---

🔗 Tech Stack

• Languages: Python, Bash (cron)
• Frameworks: FastAPI
• Libraries: spaCy, Sentence Transformers, HuggingFace Transformers, Redis-py, Pandas
• LLMs: Gemini
• Storage: Redis
• DevOps: Cron, GitHub
• Deployment Ready: REST APIs for front-end consumption

---

🧠 Sample Insights Output

• 🎯 Categories with highest volume and polarity
• ⚠️ Emerging issues and customer pain-points
• 🔍 Root cause analysis and customer intent
• 📉 Predicted business impact based on review sentiment trends

---

💼 Use Cases

• Product Teams → Understand feature-level sentiment and usability bottlenecks
• Marketing Teams → Gauge customer satisfaction and campaign resonance
• Customer Support → Detect pain points before escalation
• HR/People Ops → Analyze employee reviews from platforms like AmbitionBox

---

📌 Future Enhancements

• ✅ Multilingual support for global review processing
• 🔗 CRM integration for real-time alerting
• 📲 Expand to more review platforms (e.g., YouTube comments, Instagram DMs, etc.)

---

🖼️ Example Output & Results

• API Response...

---

🚦 How to Run This Project

1. Clone the Repository

git clone https://github.com/VedantPancholi/RevSense.git
cd RevSense

2. Install Frontend (Next.js) Dependencies

yarn install
# or
npm install

3. Start the Next.js Frontend

yarn dev
# or
npm run dev

The frontend will be available at http://localhost:3000.

4. Set Up and Run the FastAPI Backend

Navigate to the backend directory and install Python dependencies:

cd app/fastApi
pip install -r ../../requirements.txt

Run the FastAPI server:

uvicorn app:app --reload --port 8000

The backend API will be available at http://localhost:8000.

5. Configuration

• Edit .env.local and other config files as needed for your environment.

6. Access the Application

• Open your browser and go to http://localhost:3000 for the dashboard.
• The FastAPI backend endpoints will be available at http://localhost:8000.

---

📫 Contact

If you're a company looking to build intelligent feedback systems, or a team interested in using RevSense — feel free to connect.

Vedant Pancholi [LinkedIn] • [GitHub]

Mansi Thakkar [LinkedIn] • [GitHub]

Sachin Parmar [LinkedIn] • [GitHub]

Henil Prajapati [LinkedIn] • [GitHub]

---