heatdrop.ai

Classic image processing and ML algorithms implemented in JavaScript, no GPU cluster required.

Five algorithms, incremental LSI, and a CLI that actually works for multi-gigabyte datasets.

A thin, friendly wrapper around FANN that keeps the native speed while letting you stay in Ruby syntax.

A linear algebra and optimization library that stays self-contained even when doing sparse matrices, quaternions, and neural nets.

DiffSharp brings differentiable programming to .NET with a functional twist and a Torch backend.

AI-Toolbox wraps MDP and POMDP solvers in readable C++ with Python bindings, borrowing from the classic Matlab MDPToolbox and Cassandra's pomdp-solve.

A clean reference implementation of DSSM/CLSM for learning text similarity, minus the proprietary Bing data you'd need to actually run it.

A 2015-era Keras notebook that teaches RNNs to summarize news articles by learning when to paraphrase and when to just lift words from the source text.

Before WSL2 and native PyTorch wheels, getting Keras, TensorFlow, and friends running on Windows meant wrestling with Visual Studio compilers, CUDA versions, and environment variables.

A from-scratch walkthrough that treats backprop through time as a computation graph problem, not a math ritual.

A reference implementation of multilayered bidirectional LSTMs for named entity recognition, with a side of embedding archaeology.

Turns screenshots of online chess games into FEN strings using a CNN, because manually transcribing boards is even more tedious than it sounds.

A Go-native toolkit for training word embeddings when you don't want to leave the gopher ecosystem.

A Node.js wrapper for running pre-trained TensorFlow graphs without installing Python on your production servers.

A research-grade neural machine translation framework built before Transformers ate the world, still supporting RNNs, attention variants, and interactive translation.

Side-by-side exercises for developers who think in ndarrays and need to learn TensorFlow's dialect.

A bridge too far: wrapping Facebook's TorchCraft into a Gym environment for Brood War RL experiments.

A TensorFlow implementation of the Single Shot MultiBox detector trained on stop signs and pedestrian crossings, with refreshingly honest notes about its limitations.

A straightforward TensorFlow implementation of A3C that trains Pong agents for 26 hours and actually shows its work.

A readable reimplementation of guided policy search for robotic control, aimed at researchers who want to understand the machinery rather than just run it.