mimimimimi | Devpost

Inspiration

Our project was inspired by the need to transform customer feedback into actionable insights for technicians. We needed to establish faster resolutions of network issues while also recognizing regions of high customer satisfaction. T-Mobile receives feedback from calls, online reviews, community forums, and websites, but it can be difficult for technicians and management to quickly identify problem areas or recognize where the company is performing well. We wanted to create a system that automatically detects hotspots of complaints, evaluates their severity, and visualizes trends over time, making it easier for T-Mobile to prioritize fixes, improve network coverage, and enhance customer satisfaction. The idea was to combine AI, sentiment analysis, and spatial clustering to transform scattered feedback into an organized data model.

What it does

This project gathers customer feedback from various sources and analyzes it using AI and machine learning. It restructures the feedback given by the customer into a graph-based form, showing the relationships between complaints, locations, and severity. The resultant system will cluster complaints to identify hotspots, add scores by sentiment and recurrence, and visualize trends over time. Combining spatial, temporal, and textual analyses in one dashboard lets technicians and management quickly identify problematic areas, keep track of recurring issues, and make data-driven decisions for improvement in network performance and customer satisfaction.

How we built it

We developed a full-stack solution that collects, analyzes, and visualizes customer feedback to provide actionable insights for technicians and management. The front end features a dashboard with a main page showing overall trends and coverage maps for T-Mobile and competitors, a hotspots page visualizing complaint clusters and their severity, a trends page tracking temporal patterns for each region, and a log page maintaining the history of fixes. On the back end, we collect data from phone calls, speech-to-text, tone detection, website reports, and reviews up to one month old from Google Reviews, Consumer Affairs, and the T-Mobile Community Forum. We perform sentiment analysis to detect the severity of complaints and identify satisfied regions, use BERT embeddings to represent textual feedback, and apply DBSCAN clustering to detect hotspots and assign risk scores based on sentiment and cluster density. We also used an LLM (Gemini) that provides AI-driven recommendations for areas needing attention, highlights high-performing regions, and offers actionable suggestions. Additionally, coverage maps are integrated to identify T-Mobile’s strengths and weaknesses relative to competitors. The project uses Shiny Python, Geopandas, Seaborn for visualization, Next.js, HTML, CSS, JavaScript for the front end, Fast-BERT for sentiment analysis, Eleven Labs API for tone detection, MongoDB for our database, and DBSCAN for clustering, creating a real-time, AI-driven system to transform scattered customer feedback into organized, actionable insights.

Challenges we ran into

Some challenges we faced included integrating coverage maps with hotspot data, which required aligning geographic coordinates, dealing with differences in resolution between T-Mobile and competitor maps, and ensuring the data could be layered effectively without cluttering the visualization. Configuring DBSCAN to handle varying complaint densities was also difficult, as some regions had very dense clusters of complaints while others had sparse data, making it challenging to choose parameters like eps and min_samples that could accurately detect meaningful hotspots without labeling too many points as noise.

Accomplishments that we're proud of

mobling tAccomplishments we’re proud of include building a fully functional, real-time dashboard that automatically detects complaint hotspots and updates them hourly, allowing T-Mobile to quickly identify and respond to problem areas. We successfully developed a severity scoring system that combines sentiment analysis and clustering to prioritize issues, making customer feedback actionable for technicians. Additionally, we integrated AI-driven recommendations using an LLM to suggest areas for intervention and highlight regions of strong performance. We also combined coverage maps with complaint data to provide a comprehensive view of T-Mobile’s network strengths and weaknesses compared to competitors. Overall, we created a system that turns scattered customer feedback into organized, actionable insights, improving decision-making and enabling proactive network management.

What we learned

Throughout this project, we gained valuable experience in both front-end and back-end development. On the back end, we implemented machine learning algorithms such as DBSCAN for clustering complaint data, Graph Neural Networks (GNNs) for modeling spatial and relational information, and BERT embeddings for natural language understanding of customer feedback. We learned how to analyze large, multi-source datasets, combining historical statistics with real-time feedback to detect trends, recurring issues, and sentiment patterns. On the front end, we designed and implemented a user-friendly, interactive dashboard using Next.js, HTML, CSS, and JavaScript, integrating dynamic visualizations with Shiny Python, Geopandas, and Seaborn. This allowed hourly updates of hotspot maps and seamless presentation of AI-generated recommendations. Overall, we developed skills in translating complex data into actionable insights, tackling both technical challenges and usability considerations to create a dashboard that is meaningful and intuitive for technicians and management.

What's next for mimimimimi

In the future, we plan to scale this project nationally. The system will overview all regions where T-Mobile operates. This will involve handling larger datasets, integrating additional sources of customer feedback, and refining our ML models to maintain accuracy across diverse geographic areas. We also aim to enhance predictive capabilities, using AI to forecast potential hotspots before they become critical. We could use a time series model to analyze previous statistics and anticipate emerging issues. This will overall improve customer satisfaction.