AI infrastructure built on geometry,
not guesswork.
Independent research into transformer architectures, compression systems, and safety infrastructure—designed from first principles, where alignment can become a property of the system itself.
The Geometric Foundation
Our triaxial classification system maps model behavior in embedding space. Three orthogonal axes — the foundation everything else is built on.
Measures semantic proximity to harmful content archetypes. Gravity-well attractors in embedding space pull dangerous outputs toward detection thresholds.
Quantifies genuine informational value independent of harm scoring. Prevents false positives on legitimate educational, medical, and security content.
Establishes statistical normality against conversational baselines. Detects anomalous drift that single-axis systems miss entirely.
Cerberus
An experimental transformer architecture that separates local language processing, full-context reasoning, and safety-oriented Guardian attention. Cerberus v2 explores whether specialized head groups, in-forward-pass gating, curriculum training, and SNR-informed scaling can make safety constraints more structural and inspectable.
Project Portfolio
The four systems currently under active development, spanning activation-space safety sensing, transformer architecture, KV-cache compression, and non-gradient learning.
A research-preview toolkit for sensing safety-relevant signals inside transformer activation space. Includes a seven-gate research pipeline, a lightweight Sign-Check sensor, model-specific safety profiles, evaluation tooling, and experimental GGUF support.
View repository ↗A PyTorch transformer prototype separating Language, Reasoning, and Guardian head groups. The current system includes triaxial gating, hierarchical synaptic routing, iterative Wobble-Down refinement, phased training, and adaptive scaling experiments.
View repository ↗A multi-scale compression pipeline for standard transformer KV caches. It folds redundancy across layers, keyframes, and tokens, with closed-loop predictive coding, quantization experiments, Hugging Face integration, and reproducible benchmarks.
View repository ↗A CPU-native experimental learning substrate built from wave propagation, persistent scarring, sparse graphs, and prototype-bank readout—without gradients, optimizers, or backpropagation. Its current baseline was distilled from 16 falsification experiments.
View repository ↗Let's Build
Exploring novel AI architectures, structural alignment, safety infrastructure, or research collaboration? Let’s talk.