Customer Review Intelligence

Transforming Customer Feedback into Business Growth with an AI-Powered Customer Review Intelligence App

Comment

Inspiration

The project was born from a real-world business need. After rebranding a family tea business, Southern Frontier, and opening a physical store in China, we accumulated nearly 200 customer reviews from platforms like Dianping (China’s Yelp). While my team was synthesizing themes manually, my data science background demanded a more scalable and rigorous approach to convert this large volume of unstructured customer feedback into clear, actionable business insights. The recent advancements in Generative AI, especially Large Language Models (LLMs) and coding agents, made this complex ETL and NLP project feasible to build quickly.

What it does

The Customer Review Intelligence System is an end-to-end web application that automates the process of extracting, persisting, analyzing, and visualizing customer feedback. It is structured into three main layers:

  1. Ingestion (Image Processing): Uses Gemini's image recognition feature to convert raw image screenshots of Dianping reviews into structured data points (username, date, ratings, content).
  2. Persistence (Database): Stores the clean, organized customer review data in a cost-effective cloud database (Neon), complete with a cache layer for fast data operations.
  3. Intelligence (Analysis): Performs automated sentiment analysis and summarization on the complete dataset to generate actionable insights and suggestions. It also visualizes key performance indicators (KPIs) through charts (e.g., average ratings, sentiment breakdowns over time) and features a dedicated AI assistant to answer user questions about the data and reports.

How we built it

The application was built using an agent-first, vibe coding approach in Google Antigravity:

  1. High-Level Prompting: I started with a comprehensive prompt for Gemini, outlining the need for a web application to handle data ETL, visualization, and analysis (e.g., “I want to build an app to (a) take uploaded pictures... to transform into a dataset... (b) append new reviews... (c) on the front end, show consolidated data, and run a summary/sentiment analysis... Can you tell me how to build this app?”).
  2. Tech Stack Selection: Based on Gemini’s recommendations, I selected Streamlit for the frontend, Neon for the cost-effective database, and configured a Gemini API Key for the core extraction and intelligence tasks.
  3. Agentic Development: I used Google Antigravity as the coding agent to generate and implement the app’s architecture, integrating the three layers (Ingestion, Persistence, Intelligence) and significantly accelerating the 0-to-1 development process.

Challenges we ran into

The challenges were primarily in the initial data acquisition and model refinement:

  1. Data Acquisition and Evaluation: Dianping intentionally prevents web scraping, forcing the tedious task of manually taking approximately 200 screenshots. However, this served as an essential human evaluation step to validate the accuracy and relevance of the final sentiment analysis against a personal review of the feedback.
  2. Prompt Refinement: The core functionality required several rounds of prompt engineering. This involved tweaking the initial extraction prompt for Gemini to ensure accurate and consistent data capture from the diverse image screenshots and iterating on the intelligence report generation prompt to refine the output and actionable suggestions.
  3. Latency: The ETL process for batch uploading ~200 images had high latency, which was addressed through application enhancements.

Accomplishments that we're proud of

We successfully developed a sophisticated solution that provides significant business value:

  1. Multimodal Success: Successfully demonstrated the power of multimodal LLMs by reliably distilling structured, relevant data (ratings, dates, content) from unstructured image screenshots.
  2. Efficiency and Cost: Built a rigorous and scalable intelligence solution with minimal cost. The estimated cost for extracting ~200 images and running sentiment analyses was less than 1 cent per request. The MVP was functional in less than an hour.
  3. User Experience (UX) Enhancements: Implemented parallel processing for latency reduction during the batch ETL process and added a cache layer to accelerate data operations.
  4. Feature Completeness: Delivered a comprehensive app with all necessary functionalities including a virtual assistant for Q&A, a bilingual feature for reports and the ability to download data and reports for external analysis (e.g., importing into NotebookLM to create a slide deck).

What we learned

The project served as a powerful validation of the shift in data science and application development:

  1. Accessibility of Advanced Tools: Generative AI and modern LLMs have fundamentally changed the landscape, simplifying complex tasks like data extraction from unstructured image data and automating sophisticated sentiment analysis.
  2. Agentic Speed: The process validated that even complex ETL and NLP pipelines can be rapidly built using accessible AI coding agents with minimal time and cost, making advanced data projects accessible to small business owners.

What's next for Customer Review Intelligence

The immediate future for the application is to focus on providing continuous, actionable feedback to drive ongoing product and service improvements for Southern Frontier. Potential next steps include:

  1. New Data Sources: Expanding the ingestion layer to pull data from other review platforms for richer, multi-dimensional analysis.
  2. Customization: Developing more advanced customization options for visualization and report generation based on evolving business needs.

Built With

Share this project:

Updates