NeuralNet
Visual Deep Learning Architecture Studio & Code Compiler
NeuralNet is a full-stack visual deep learning studio where you design neural networks with drag-and-drop nodes, compile to Keras, train with real TensorFlow metrics, and deploy secure prediction APIs.
What Jayant Potdar Built & Owned
Led the project from conception to working prototype: created the interactive node-graph canvas using React Flow, engineered the topological sorting graph compiler in FastAPI that converts node JSON into valid TensorFlow/Keras Python scripts, and built the WebSocket-backed training monitoring interface that streams live epoch loss curves.
Key Technical Decisions:
- Topological sort compiler converting arbitrary directed acyclic graphs into clean, PEP-8 compliant Keras code.
- Real-time tensor shape validation catching dimensional mismatches before training begins.
- WebSocket streaming of training metrics (loss, accuracy, validation loss) rendered with live charts.
- One-click export of saved `.keras` models accompanied by an auto-generated FastAPI inference endpoint.
Core System Features & Capabilities:
- ✓Drag-and-drop layer nodes and connect them to design a model visually.
- ✓Converts the visual graph into executable TensorFlow/Keras model code.
- ✓Supports built-in datasets like MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and custom dataset uploads.
- ✓Lets users configure training and view metrics such as loss and accuracy over epochs in real time.
- ✓Trained models can be deployed as production API endpoints with generated API keys.
- ✓Provides an in-browser UI to test deployed models by sending inference requests and viewing predictions.
Translating a visual DAG (directed acyclic graph) into valid sequential or functional Keras code requires strict validation: cyclic loops must be blocked, branching tensor dimensions must align, and layer parameters must typecheck.
I implemented client-side connection constraints in XYFlow, and built a backend validation engine that performs topological sorting on the graph JSON. The compiler constructs a clean Keras script, executes local training in an isolated process, and pushes real-time epoch statistics over WebSockets.
Detailed interface walkthrough, diagnostic panels, and feature screenshots for NeuralNet.
The working prototype lets users compose Dense, Conv2D, MaxPooling, and Dropout layers visually, train models on standard classification datasets, and instantly generate a production-ready Python inference endpoint.
Building a Visual Deep Learning Studio: Compiling Graph JSON to Executable Keras Code
Transforming interactive React Flow DAGs into topologically sorted TensorFlow models with live WebSocket training.
Read the engineering note (8 min read)
AstroPixel
AstroPixel is a FastAPI + React platform for exploring NASA gigapixel imagery with buttery-smooth deep zoom, annotations, and secure user/admin workflows. It ingests GeoTIFF/PSB files, generates tile pyramids with GDAL, and serves them through an optimized viewer.
Krypton
Krypton is an AI-powered research assistant that searches arXiv and OpenAlex, ranks papers by relevance, recency, and citations, and helps users understand literature faster with Gemini-generated summaries, structured insights, research-gap analysis, and personalized topic recommendations in a FastAPI + React app.
