Currently playing around with AI systems that can remember, reason, and occasionally justify the confidence.
Memory, retrieval, evaluation, and world models are the recurring pieces. I care about systems that use evidence, survive long horizons, and stay legible when they fail, which is usually where the interesting part starts.
This is a small public shelf: runnable references, research scaffolds, and notes that escaped the notebook.
Current public work: nanorlm.
Legacy: Cnn-Hypercolumn · Jupyter-Notebook-Atom-Scheme.