LitPath Labs

AI Systems & Prototyping

Targeting model obsolescence through user self-sufficiency. We build strictly for technical efficiency, empirical rigor, and uncompromised accuracy.

Active Research & Deployments

Contextual Norm Propagation

Active · re-audited September 2026

A controlled multi-agent financial market simulation evaluating accountability under hidden risk. The September 2026 audit reproduced the scores of all 20,520 saved agent turns; when accountable agents published their notes, accountable language rose against a silent control (+0.256, 95% CI 0.229–0.283) and survived ablation. Speaking order confounds exposure, so the next round randomizes it.

LLMs Multi-Agent Systems Reproducibility

Local Inference on GB10

Active · in daily use

Two NVIDIA GB10 nodes (128 GB unified memory each) replace hosted models. One serves a 125B-parameter mixture-of-experts model (6B active) through llama.cpp at 23.6 tokens/s with 0.33 s to first token; the other runs image, video and 3D generation. Every service runs on owned hardware with owned weights.

GB10 llama.cpp MoE

Aethelgard's Respite

Active

A persistent AI layer for the Creation Engine. NPC dialogue and memory run in-process through SKSE and SkyrimNet, served by a fast local model on the GB10 nodes; a separate chronicler reads in-game memories and turns them into illustrated journals.

SKSE Python Game Engine APIs

Picture-to-Game Asset Pipeline

Active

A written description becomes a game-ready object: concept art, a 3D mesh (Hunyuan3D 2.1), a skeleton (UniRig), animation from a motion library, then delivery to Godot. An unattended factory queue builds whole catalogues of props for interactive projects.

3D Generation Blender Godot

Scalable Hydroponic Infrastructure

Active

Engineering closed-loop controlled environment agriculture (CEA). Prototyping focuses on optimizing system fluid dynamics and integrating sensor telemetry for automated nutrient and environmental regulation.

Hardware Fluid Dynamics Telemetry

Writing

Without Lying to Yourself

First edition · free to read

A field manual for running AI workloads on GB10 hardware, built from real incidents: twelve plates on the failures that look like success — a healthy ping on a wedged box, an idle process squatting on memory, a job that exits clean and delivers nothing.

Field Manual GB10 Operations