Neural / AI
Self-hosted deep learning.
A tensor reverse-mode autograd engine written in C++, with a PyTorch-style layer API. Transformers, LSTM and CNN, deep reinforcement learning, anomaly detection, and a self-hosted language model. No Python, nothing leaves your machines.
The foundation
One autograd engine, every architecture.
Nodes are tensors, ops carry a tensor-level backward, and a dynamic tape walks the graph in reverse. Add an op once and every model that uses it works. The whole point: every op's analytic gradient is checked against a finite-difference gradient, so the math is proven, not assumed.
- Double-precision, deterministic seeds
- Eigen-backed matmul, SIMD, OpenMP
- PyTorch-style Module / Layer API
- Train, save, load, predict, zero deps
$ ./nn_recovery
[PASS] gradient check (matmul) |d|~1e-10
[PASS] gradient check (attention)|d|~1e-10
[PASS] MLP learns XOR
[PASS] CNN classifies bars 100%
[PASS] LSTM recall (BPTT) >90%
[PASS] Transformer next-token 9/9
72 checks PASS
What is inside
A complete stack, in one library.
Layers
Dense, Conv1D/2D and pooling, RNN, LSTM, GRU, multi-head attention and full transformer blocks, embeddings, batch / layer / group norm, residual.
Training
Adam, AdamW, RMSprop, and more, with LR schedules, grad-clip, early stopping, checkpointing, and metrics. Mini-batch DataLoader.
Anomaly detection
Autoencoder and VAE, LSTM-autoencoder, deep one-class, isolation forest, and Mahalanobis, with reconstruction scoring and thresholds.
Reinforcement learning
Tabular Q-learning, SARSA, and double-Q, then DQN (double, dueling, replay, target net) and policy gradient (REINFORCE, A2C, PPO) on gridworld and cartpole.
Self-hosted LLM
A character transformer you train, save, load, and generate from with zero dependencies. Loss falls 4.36 to 0.0003 and it regenerates the corpus exactly.
Dimensionality
PCA and SVD utilities, plus the learned latent spaces of the autoencoders and the transformer encoder.
Proven on real data
ROC-AUC 0.95 on real card fraud.
On the ULB credit-card dataset, 284,807 real transactions with a 0.17% fraud rate, the unsupervised anomaly stack scores ROC-AUC 0.95 with no labels, the published benchmark. The same stack flags faults, intrusions, and quality defects in other sectors.
ROC-AUC, unsupervised, real fraud
real transactions scored
starter demo, train to result
Python, it is all C++
Editions
Start small, or take the whole suite.
AI Starter
Autograd with gradient-check, MLP and CNN, optimizers and schedules, the training loop, PCA, and a fast demo.
AI Pro
Everything in Starter, plus RNN / LSTM / GRU, transformers, deep RL, the anomaly and fraud stack, and the self-hosted LLM.
Run it on your own hardware.
No API, no per-token bill, no data leaving your network. We will show the gradient-checks and the fraud demo live.