Builder Projects
Ship, Don't Theorize.
Open-source tools and resources from The New Guard.
active
TNG LLM Bench
Five-round local-vs-frontier model bake-off on one RTX 5090. Fully reproducible.
PythonGoShellllama.cppPiClaude Code
Companion repo for the Five Rounds Deep Dive. Every brief, per-model implementation, evaluation script, log, and screenshot from a five-round bench across four open-weight 30B-class locals plus Codex and Claude Sonnet 4.6. Runs in under an hour on the same hardware.
View Project → active
tng-wiki
Keep your LLM wiki honest
Node.jsShell
A per-claim citation lockfile, drift detection, and code-grounding for LLM-maintained knowledge bases. Scaffold a Karpathy-style wiki in one command, then keep it trustworthy as it grows instead of letting it silently rot.
View Project → experimental
tng-heretic
Reproduce our Gemma 4 abliteration experiment
PythonCUDANode.js
Companion repo for the Gemma 4 abliteration lab. Everything needed to reproduce the 24-minute Heretic run against Gemma 4 E4B-it on a single GPU, plus the comparison tooling and results.
View Project → active
TNG Toolbox
Developer tools and workflow extensions for Claude Code
PythonShellClaude Code
A Claude Code plugin marketplace — install individual tools for git provenance, code review, testing, and the glue between AI coding sessions and your existing workflow.
View Project → archived
NemoClaw Quickstart
Ship your first secure AI agent — sandboxed, policy-governed, GPU-ready
PythonDockerk3sNVIDIAWSL2
Workaround repo for deploying sandboxed OpenClaw agents with early NemoClaw releases. Handled the WSL2 setup the official installer couldn't at the time. Archived now that NVIDIA's tooling has caught up; kept for reference.
View Project →