DoubleCloud with GPT-4

Sub-second analytics platform on open-source technologies

Analytics

Data & Analytics

Business Intelligence

Visit WebsiteSee on Product Hunt

Hey PH, Victor is here, A few months ago, here at DoubleCloud we discovered that GPT-4 could understand and analyze structured data like tables, HTML, CSV, and other formats. We have a self-service BI at https://double.cloud where users can create vibrant and appealing charts and dashboards. We often hear from users who struggle to understand the information presented on dashboards created by their analytics teams. They face challenges interpreting a variety of puzzling KPIs, metrics, complex graphs, etc. This is why we built in an integration with GPT-4: to help end users quickly get explanations, in a human-readable format, on what's happening in a particular chart. GPT-4 provides conclusions, insights, and even recommendations in that specific domain of knowledge. We are creating a prompt behind the scenes that helps the AI understand the context of the chart and pass the raw data of that chart to it. This tool becomes extremely valuable for complicated charts and tables where the human brain needs time to understand the context, the numbers, scale, and other different factors. It's like having a business analyst who can quickly give you an overview and business advice. Here are some other examples: 1. When analyzing data from marketing campaigns, it may reveal trends about advertising channels’ efficiency and audience behavior. This, along with recommendations to improve metrics like CAC, ROMI, etc. makes it an invaluable tool for business owners. 2. Another real example is in the food delivery area, where the AI automatically understands the business context and identifies anomalies across different warehouses based on their KPIs. Given that the number of KPIs was about ~20 across roughly 30 warehouses, it's not easy for humans to quickly understand and compare so many data points. However, GPT-4 identified a few warehouses with significantly low speed during the packaging stage and suggested moving staff from these warehouses to a particular warehouse to learn from the best one. As for future ideas, we are waiting to gain access to the multimodal version of GPT-4 to test and potentially implement the idea of using images, perhaps passing the full screenshot of a dashboard with different charts. This could give more context to the model and potentially enable it to find correlations between charts, providing broader and richer answers. These answers could not only explain a single chart but also provide an overview of the state of your business based on the range of the dashboards. The main limitation to implementing this with raw data is the limit of prompt length, and data in a visual format could be more compact/compressed.