My approach was to build a lightweight, AI-powered tool to help a small business understand sentiments and themes in their Google Reviews. I started with a Streamlit app as the user interface, integrating Hugging Face’s DistilBERT model for real-time sentiment analysis. To store review embeddings efficiently, I used Pinecone as a vector database, enabling future retrieval and analysis. Zapier was employed to automate data logging into Google Sheets. To meet the problem statement’s requirements, I enhanced the app with BERTopic for topic extraction, Plotly for time-series sentiment trends, and a keyword-based classifier for categorizing reviews by service, location, or product. A weekly summary report was added using Zapier’s scheduling features.
- Streamlit: Chosen for its simplicity in creating an interactive web app, ideal for real-time user input and visualization.
- Hugging Face: Utilized the
distilbert-base-uncased-finetuned-sst-2-englishmodel for accurate sentiment classification, leveraging its free-tier API. - Pinecone: Selected as a vector database to store 384-dimensional embeddings from
sentence-transformers/all-MiniLM-L6-v2, supporting scalable review storage within the 5GB free tier. - Zapier: Integrated to automate data transfer to Google Sheets and schedule weekly email summaries, fitting the 100 tasks/month free limit.
- BERTopic and Plotly: Added for topic modeling and time-series visualization, enhancing analytical depth with open-source tools.
- Python Libraries (Pandas, Requests): Used for data handling and API calls, ensuring a robust backend.
These tools were picked for their accessibility, free-tier compatibility, and ability to integrate into a cohesive workflow.
Several challenges arose during development:
- Hugging Face API Downtime: Frequent 503 errors required adding retry logic, falling back to a “Neutral” sentiment after three attempts.
- Pinecone Region Restriction: The free tier didn’t support
us-west-2, necessitating a switch tous-east-1and recreation of the index. - Zapier Runtime Limits: The lack of
sentence-transformersandpineconein Zapier’s environment forced me to move embedding generation and upsert operations to Streamlit. - BERTopic Compatibility: An
ImportErrorforStaticEmbeddingdue to version mismatches betweensentence-transformersandbertopicrequired downgrading to compatible versions (e.g.,sentence-transformers==2.2.2,bertopic==0.15.0). - Time Constraints: Fitting topic modeling, time-series charts, classification, and weekly reports into the deadline limited thorough testing.
Hands-on experimentation helped overcome these by adjusting configurations and leveraging documentation.
With additional time, I would:
- Enhance Topic Modeling: Train a custom BERTopic model with more reviews to improve topic coherence, potentially using LangChain for advanced NLP summarization.
- Refine Classification: Replace keyword-based classification with a machine learning model (e.g., a multi-label classifier) trained on labeled review data for service, location, or product categories.
- Automate Reports: Develop a Python script to generate detailed PDF reports with charts, triggered weekly via Zapier, instead of basic email digests.
- Add User Feedback: Implement a feedback loop in Streamlit to allow business owners to flag misclassified reviews, improving model accuracy over time.
- Scale Pinecone: Explore paid tiers to handle larger datasets and multi-index setups for different business locations.
These enhancements would make the tool more robust and tailored to a multi-unit business, aligning with the problem statement’s full scope.
- Length: Fits within 1 page (approximately 250–300 words).
- Structure: Follows the required sections: approach, tools, challenges, and future steps.
- Customization: Replace placeholders (e.g., specific version numbers, exact error details) with your actual experience if needed.
- Next Steps: Share this write-up along with the Loom video link, Google Sheet link, Pinecone stats, and Zapier Task History after recording the video.