Daniel

Unified agent memory and context compression stack for 2026 NVIDIA + edge (Vera CPU, Grace, Jetson Thor, 3090). Glues busyBee-cpu, honey-comb, and rust-brain. Better effective reasoning per token.

Rust CLI for RBMEM (.rbmem), a structured Rust-Brain memory format with timestamp-protected sections, hierarchy, graph relations, Hermes integration, and Markdown sync.

No description.

A lightweight CPU-only classifier that cuts AI API costs in agentic coding workflows by routing simple decisions (read file, run tests, apply patch) to a local 30ms model instead of calling a cloud language model, only escalating when real reasoning is needed.

No description.

CPU-only inline context compression for agent harnesses. Keep the honey, drop the wax.

Open-source toolkit for mapping and visualizing Mixture-of-Experts model routing patterns

A 2MB sklearn classifier on consumer CPU beats Cohere's $10M+ 218B MoE on HumanEval

A Python tool that runs AI coding benchmarks locally on a consumer CPU using a 2.1GB quantized model, demonstrating that a small Qwen coding model can score 89% on HumanEval and beat larger commercial models on narrow coding tasks.

No description.

Spectral Analysis-Based Entanglement Resolution — ablation pipeline for removing refusal directions from LLMs

llama.cpp fork with a patch for RYS-duplicated Qwen3.5/Qwen3Next models (non-uniform full_attention pattern). See rys-qwen35 branch.

Talos trace-curator: Hermes Skill that turns raw agent sessions into training datasets

A reusable, model-agnostic methodology for evaluating pruned/compressed Mixture-of-Experts (MoE) model releases: rubric, red-flag catalog, runnable behavioral tests, scorecard + producer checklists, and worked examples.

Tower defense game with a Gymnasium API for training AI agents (RL, heuristic, tournament).

Bio-inspired CNN modeling M/P/K pathways of the Lateral Geniculate Nucleus. From 76KB/15.5K params (33 FPS on Raspberry Pi) to full-scale and temporal variants. Matches ResNet18 with 52x fewer parameters. No pretraining required.

T013: Reasoning Depth Assessment - Logical Fallacies. A psychometric benchmark for evaluating AI agent reasoning quality using Item Response Theory (IRT) principles.

axoFast — make Axolotl QLoRA fine-tuning fast: the Unsloth-equivalent kernel recipe (flash-attn2 + Liger), head-to-head benchmarks vs Unsloth, and the fixes to run it on the NVIDIA DGX Spark (GB10/sm_121) and on x86 cloud GPUs.

AI Research Portfolio — Dan Lougen

AI Research Portfolio — Dan Lougen

Native Rust port of the Hive agent memory + context compression stack. 31 tests, 240k msg/s on x86_64, Android + Jetson cross-builds.

Kakeya polynomial partitioning for GPU ray tracing load balancing

Kakeya polynomial partitioning for GPU tile workload balancing in 3D Gaussian Splatting

Play Risk in the browser — Three.js 3D board, AI opponents (heuristic or LLM), full game rules

Classic space invaders game - playable directly in browser. Built and one-shotted by Ornstein-122B-A10B

No description.

The Flight Simulator for Production AI Agents — Generate high-quality synthetic trajectories for training reliable SRE, DevOps, and infra agents.

Experiments coded by myself for current projects.

The agent that grows with you

Caduceus T014: Build a 5-page SaaS website from a locked golden spec. Tests frontend skills, accessibility, interactivity, and spec fidelity.

No description.

personal webstie

No description.

Second-order motion Inhibition of Return experiment using texture-defined stimuli

Adaptive staircase experiment to find the IOR (Inhibition of Return) crossover point using jsPsych

No description.