Neural Network Trainer
The problem
The project started from a real frustration on our own team. One teammate works in machine learning, and every time he wanted to train a model he was hand-writing Python scripts from scratch — editing code just to change a layer or an epoch count, re-running from the terminal, with no reusable interface and nothing a non-coder could touch.
The solution
Neural Network Trainer turns that workflow into a point-and-click tool — no code, no notebooks. You drag in a CSV, pick the target column and features, stack layers (Dense, Dropout, Conv1D, LSTM…), set epochs and batch size, and start training. A live progress view polls the training loop and charts loss and accuracy as they update; when it's done you can download the model or test it in an interactive playground.
Under the hood, training runs in C# through TensorFlow.NET — built, compiled, and fit entirely inside one .NET app, with EF Core / SQLite persisting each run. Built by a three-person team for a .NET course; I owned the frontend and the data layer, while the ML teammate built the core training engine (see my contributions below).
Key features & challenges
- Point-and-click neural-network training in the browser — drag a CSV, stack layers, hit train.
- Live training metrics: the UI polls the training loop and charts loss/accuracy per epoch.
- Training runs in C# via TensorFlow.NET — no Python, all inside a single .NET app.
- Trained models are downloadable and testable in an interactive playground.
My contributions
- Built the Blazor frontend: drag-and-drop CSV upload, dataset preview, dynamic layer/parameter configuration, and the live training-progress view.
- Built the data layer: EF Core + SQLite persistence, CSV parsing, and the training-form model.
- Contributed parts of the REST API (training start / status endpoints) that the UI polls for live metrics.
- Core C# TensorFlow.NET training engine was built by a teammate (three-person team).