Systems Under Construction

Measure the savings. Measure the loss. Keep the result either way.

A falsifiable MLX research program testing whether structured tensor-axis operators can make language-model training more resource-efficient on edge hardware without losing the quality frontier.

The source becomes the glyph field.

A native macOS creative application that converts photos and video into fully glyph-rendered visual media. Metal analyzes the source and composes the output field entirely from glyphs, then exports it as stills, audio-preserving video, or ANSI / HTML / text. The same Metal 3 engine — continuous-coordinate simulation, per-cell GPU compute, temporal stabilization — also drives real-time rendered scenes.

H-E-R blocks: recurrence, sparse attention, and expert routing in one stack.

A family of hybrid recurrent language models — HarmonicStateLayer, ElasticAttentionGate, and ResonantExpertRouting blocks — trained entirely on Apple Silicon.

The D-CSIL-1/2/3 research and training line.

The D-CSIL-1/2/3 training line: selective state-space and GLA models pretrained on all-local pipelines.

Not yet a validated breakthrough — and it says so on the blueprint.

Brain-inspired sequence models: O(L) recurrent cores, Hebbian plasticity phases, neuromodulation, and sleep-style consolidation — now converging on the HibbieFormer-HX program.

The kernels that make local training fast enough to matter.

Fused Metal GPU kernels for linear recurrence on Apple Silicon: Mamba-2 selective scan, gated linear attention, RG-LRU, and rotational LRU — with full training support.

Your AI research partners never sleep.

The lab's agent program: durable, terminal-first agent runtimes with a deterministic policy engine, reversible git-backed edits, budget-aware model routing, and Cortex shared memory. Remy leads on a cloud model; Justin assists from a second instance on a local one.

One memory, many agents.

Durable shared memory (Cortex) and disciplined session orchestration (Harness) for every agent in the lab, served from a Raspberry Pi 5 named Bailey.

Claims must survive the Skeptic to become results.

A governed research loop: objective → plan → agents → experiments → evidence → critique → revision, with claim gates and immutable audit logs.

Structure as a substitute for scale.

Councils, hives, and debate: multiple specialized models coordinating on one problem — from a Raspberry Pi MQTT hive to a five-model local debate framework.

Generate it, separate it, say it — locally.

Machine learning for sound: conditional drum generation, 6-stem source separation on MLX, XTTS voice cloning, and edge TTS.

What if regularization looked like field dynamics?

A physics-inspired architecture: multi-scale field layers with flux, Schwarzian, and latent-diffusion regularizers over causal attention.

A cool idea — still on paper, and labeled that way.

A design for computational interpretation of canine communication from audio, video, pose, and interaction context.

Build it, break it, measure what happens, redesign until it works.

Electronics prototyping and sensor systems that connect intelligence to the physical world — from an RV leveler to machine-health monitoring and the lab's embedded fleet.