Predictra

Home Page
Model Training Page
Training Results Page
Analysis and Chatbot Page 1
Analysis and Chatbot Page 2

Inspiration 💡

We were inspired to build Predictra by the frustrating gap between the vast amount of available data and the valuable predictive insights that often go untapped. Traditional machine learning is a months-long endeavor requiring deep expertise, while existing no-code tools are often too simplistic and lack the power of modern AI. Predictra was born from a single question: "What if we could put the power of a custom PyTorch neural network into the hands of anyone with a CSV?" We built a platform to be that bridge, combining a simple drag-and-drop interface with the sophistication of real-time model training, all visualized live via WebSockets. Our inspiration is to finally democratize data science, empowering anyone to turn their raw data into actionable predictions in minutes, not months.

What it does 🔮

Predictra is an intelligent data analysis platform that transforms raw CSV datasets into powerful, actionable insights. It completely brings machine learning to anyone, allowing people to harness the power of predictive analytics without writing a single line of code. Users simply drag-and-drop their own CSV file, and Predictra's AI automatically analyzes the data, identifies data types, and prepares it for modeling. From there, users can select a target variable and instantly begin training a dynamic neural network. The platform features a real-time, WebSocket-powered dashboard where users can literally watch their model learn, and once trained, they can immediately use an interactive form to make live predictions on new data. Predictra is the bridge from raw data to real predictions, in minutes.

How we built it 🛠️

We built Predictra on a modern, high-performance tech stack, starting with a FastAPI (Python) backend to serve a fast, asynchronous RESTful API for all data and model management. The core of our "Zero-Code" AI is a PyTorch-based dynamic neural network, combined with an intelligent scikit-learn preprocessing pipeline that automatically handles feature encoding and scaling. The entire user experience is a fluid, responsive single-page application built in React, which uses Chart.js to render all statistical analyses and visualizations. Furthermore, we use WebSockets, which create a live, persistent connection between the FastAPI backend and the React frontend, allowing us to stream training loss metrics directly from the PyTorch model to the user's dashboard in real-time.

Challenges we ran into ⚠️

Our most significant challenge was the steep learning curve on the frontend. Two of our key team members had very little experience with React before this hackathon. They had to learn the entire framework from components and state management to hooks on the fly, all while simultaneously building our complex, responsive UI, including the drag-and-drop uploader and dynamic charting. This was compounded by our most ambitious technical feature: the real-time training visualization. Getting the FastAPI backend, the PyTorch model (running in its own process), and the React frontend to communicate seamlessly via WebSockets was incredibly difficult. We spent a significant amount of time debugging the data pipeline to get the loss metrics to stream from the model to the API and then be broadcast instantly to the user's Chart.js graph. Making this update live without lagging or breaking the React state was a major hurdle that took a lot of perseverance to overcome.

Accomplishments that we're proud of 👑

We are immensely proud of delivering a polished, production-ready platform that feels like a true product, not just a prototype. A massive accomplishment was building this fluid and modern UI, as two of our team members learned React entirely from scratch during the hackathon to make it happen. Our biggest technical triumph is the "live learning" dashboard; we successfully engineered a complex pipeline using WebSockets to stream metrics directly from our PyTorch model, allowing users to watch the AI train in real-time on a Chart.js graph. Finally, we're proud of building a truly dynamic and intelligent "Zero-Code" system. Our platform automatically analyzes any uploaded CSV, adapts the neural network's architecture, and intelligently preprocesses all features, making advanced predictive analytics genuinely accessible to anyone.

What we learned 📝

This hackathon was in full-stack integration and rapid skill acquisition. We learned firsthand how to architect a complex, real-time application, successfully weaving together a FastAPI backend, a PyTorch model, and a React frontend using WebSockets, a feat that taught us how vital live feedback is for building user trust. Our biggest technical lesson was that a "Zero-Code" platform's power isn't just in the model, but in its intelligent preprocessing pipeline; handling the complexities of automatically cleaning, encoding, and scaling any user-uploaded dataset proved to be the most critical challenge. Above all, with two team members learning React from scratch to build our polished UI, we learned that our team's capacity to adapt and master new technologies under pressure is our greatest asset.

What's next for Predictra ➡️

Our next step for Predictra is to take our MVP to production by containerizing it with Docker and deploying it on a scalable cloud platform like AWS or Google Cloud. Once live, we will rapidly expand our AI capabilities beyond a single neural network, introducing a library of diverse ML models like Gradient Boosting and Random Forest to handle a wider range of problems. We will also enhance the user experience with industry-specific templates for common use cases like sales forecasting and churn analysis. Finally, our long-term vision is to build out the full SaaS business model, launching our subscription tiers, developing a robust API for enterprise integrations, and adding collaborative team workspaces to make Predictra the go-to platform for data-driven teams.