machine learning

Research codebase exploring hierarchical predictive coding as a training principle for transformers (HERT and variants).

Subproject of Playbooks: fine-tune a small open-weight model (Qwen-3-4B / Gemma) to compile and interpret Playbooks programs locally.

Karpathy-style nanoGPT fork exploring adaptive-compute transformers with per-block evidence accumulation.

Pattern-discovery engine - long-running personal research on spiking-NN pattern formation.

Neuron models that fire more readily given the right top-down context - a mechanism for mixing prediction with signal.

Training a recurrent network to settle into its attractor state faster - LISSOM, but differentiable.

BERTRAND-DR - a published discriminative re-ranker that improves neural text-to-SQL parsers, reaching top-4 on the Spider benchmark.

Learning latent-variable probability distributions directly, as histograms over quantized bins.

Fine-tunes BERT for token-level sequence tagging on natural-language utterances.

What if a variational autoencoder were used as a convolution kernel?

A series of research projects towards biologically inspired AGI.

Information-density normalization - adjust histogram bins so precision follows where the data is.

Notebook-driven exploration of free-energy minimization for perception, in pure NumPy.