-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Knowledge base for all sw30labs repositories, organized by domain. Maintained using the LLM Wiki pattern — a hand-editable markdown knowledge base as a disciplined alternative to RAG.
Author: Nicolas Cravino | Repos: 40 live (25 archived repos are not tracked) | Articles: 34 | Last updated: 2026-08-30
LLM integrity testing, regulatory intelligence for global pentest compliance, autonomous pentesting research, and enterprise AI-assisted pentesting specifications.
| Repository | Description |
|---|---|
| tslit-dspy-dgx | TSLIT — time-shift integrity testing on DGX Spark (Ollama, Muse-light detective; Qwen is a scan target, never the detector). Lineage: tslit |
| pentest-regulatory-intel | RegIntel — AI-powered pentest regulation inventory across 20+ jurisdictions with reflection quality gates |
| strixresearch | Research and documentation for the Strix autonomous AI pentesting platform |
| agentic-ai-pentesting | Book companion — two approaches to AI-assisted pentesting (autonomous platform + Burp Suite co-pilot) |
| skillspector-trial | Skillspector — single-file offline scanner grading Agent Skills A–F on security + quality, now with an optional loopback oMLX "AI Analyst" that triages, adjudicates, and issues an install verdict (static evidence stream for oscal-skills-guardrails) |
| strix-omlx | Points the Strix autonomous pentest agent at a local OMLX MLX server (abliterated MiniMax-M2) — fully local, LiteLLM routing; companion to strixresearch |
| STRIDE-Lite | Local STRIDE/DREAD threat models, ATT&CK-templated kill-chain scenarios, and a linked-note Vault with a phase × lane Campaign Score — LangGraph, OpenAI-compatible or local oMLX |
| LangGraph_STRIDE | Dual-platform LangGraph STRIDE/DREAD threat models plus CVE/CTI/scenario graphs — same codebase on DGX Spark (vLLM) and Apple Silicon (oMLX); Maintain studio + optional Neo4j |
Category page: ai-security-pentesting · TSLIT lineage: tslit
NIST OSCAL-powered tools for agent guardrails, digital twin compliance, Zero Trust posture analysis, and compliance-as-code workflows.
| Repository | Description |
|---|---|
| oscal-agent-guardrails | OSCAL profiles as a policy brain to guardrail LLM agents at runtime (allow/deny/needs_approval) |
| oscal-digital-twin-playground | Digital twin drift detection — SSP vs live config, with risk assessment and mitigation |
| oscal-zero-trust-lens | Zero Trust coverage analysis across 7 dimensions from SP 800-53 controls |
| oscal-agent-lab | Multi-agent lab — RAG Q&A, SSP diff, profile generation, validation over 1,196 controls |
| oscal-cac-playgd | Compliance-as-code CLI — explain OSCAL files, suggest remediation, PR-style diff review |
| genai-regulatory-intel | RegIntel-AgenticAI — autonomous GenAI regulatory intelligence across 16 global financial jurisdictions (LangGraph 4-agent state machine) |
| oscal-skills-guardrails | OSCAL-as-policy for Agent Skills — dual-evidence admission (static scan + local LLM rubric judge), digest integrity, CI gate, assessment-results audit trail |
| driftlab-mlx | Runtime compliance observer — diffs a live agent's decision traces against a certified OSCAL baseline, proves drift via sandboxed micro-experiments, emits OSCAL 1.1.2 assessment-results (CA-7); v2.1 adds DL-8 certified resource budgets (renamed from driftlab) |
| driftlab-dgx | DriftLab on NVIDIA DGX Spark — same 2.1.0 core and deterministic compliance path, advisory layer on the host's shared local vLLM stack |
| oscal-presence-gate | OSCAL as the policy brain for WHERE, not just WHAT — presence verifier ahead of the enforcer for delegated agent traffic (EO 14117 / DOJ DSP); unknown presence fails closed; first public OSCAL encoding of the CISA Security Requirements + 5-control PRES overlay |
Category page: oscal-compliance
Design patterns, orchestration, workflow conversion, coding assistants, and research automation using LangGraph.
| Repository | Description |
|---|---|
| agent-stack | Interactive 10-layer architecture visualization for reliable AI agents (vis.js) |
| deepagent-azure-cli | Turnkey coding assistant CLI — Azure OpenAI + LangChain DeepAgents, Textual TUI, HITL |
| ralph-dgx | DeepAgents Code CLI + Ralph goal loop on DGX Spark — install→patch→overlay harness against local vLLM (Qwen3-Coder-Next-FP8) |
| wiki-vs-rag | Four-arm benchmark over the sw30labs wiki — single-shot RAG vs agentic-RAG vs wiki-nav vs QMD; agentic-RAG wins Pareto |
| langgraph-checkpoints-vs-stores | Runnable offline reference — thread-scoped checkpoints vs cross-thread stores, real StateGraph/InMemorySaver/InMemoryStore, CI-gated; production backends (SQLite/Postgres/Redis) + HITL/time-travel chapters |
| loopscope | Local live dashboard for LangGraph graphs and Ralph loops — in-process hook, :7788, ring buffer over a websocket, JSONL record/replay; forget the hook and the agent still runs |
| singularity-atlas | The Singularity Atlas — fuses public AI-buildout feeds into a globe, eight vector panels, a 0–100 Singularity Index and a locally-written daily brief (LangGraph → Neo4j, optional Ollama) |
Category page: agentic-frameworks · Atlas: singularity-atlas
On-device AI toolkit for Apple Silicon — inference serving, benchmarking, distillation, vision-language, TTS, STT.
| Repository | Description |
|---|---|
| tars-ai | TARS from Interstellar as a local voice agent — LLM + TTS served by a local OMLX server (OpenAI protocol), zero cloud; LangGraph dropped July 2026 for a plain orchestrator, now with a multi-character crew and a TUI |
| screen-lens-mlx | Video scene intelligence — hybrid keyframe detection, local VLM captioning, ChromaDB search, code/docs/demo reconstruction. Re-converged to dual-platform in Aug 2026: one client over vLLM/CUDA or oMLX/MPS, split vision/text model roles, web command deck |
| mlx-distillation-explained | Educational distillation PoC — Claude Sonnet → Llama 3.1 8B via LoRA on Apple Silicon |
| mlx-responses-api-server | OpenAI/Azure/Anthropic-compatible local inference server with tool calling (renamed from local-mlx-responsesAPI-server) |
| QWEN3-VL-Python-OCR-Script-MLX | Batch image captioning with Qwen3-VL-30B on MLX |
| MLX-YouTubeScribe | YouTube transcription using local Whisper models with Streamlit UI |
| sulphur-2-base | Local MLX video generation wrapper for Sulphur 2 via ltx-2-mlx runtime |
| supertonic-3-mlx | Local MLX TTS for Supertonic 3 — JSON graph topology + NPZ weights |
Category page: local-inference-mlx
CUDA counterpart to the MLX toolkit — local inference on NVIDIA DGX Spark (GB10, Linux aarch64) via local vLLM and CUDA llama.cpp.
| Repository | Description |
|---|---|
| screen-lens-dgx | DGX-only ScreenLens fork — vLLM (Qwen3.6-27B-FP8) captioning, OpenCLIP on CUDA, ChromaDB, Docker compose path |
| AI-OS-1 | Research corpus (no runtime) for a model-aware distributed inference OS on N ≥ 2 DGX Sparks — paper, frozen schemas, append-only design forum |
Category page: local-inference-dgx
CLI utilities, code intelligence, and infrastructure for managing repository fleets.
| Repository | Description |
|---|---|
| gitnexus_fleet | Clone, index (KuzuDB graph), and query entire GitHub orgs via MCP + web dashboard |
| RepoBundle | Export/import Git repos as single human-readable text files |
| cleanroom | Agentic file-hygiene pipeline — Analyze → Recommend → Remediate → QA over one file via LangGraph, oMLX or DGX vLLM, honest about what it can and can't verify |
Category page: developer-tools
One-off research instruments that do not sit in security, OSCAL, agents, or local inference. Two encoding instruments: Mac and Spark.
| Repository | Description |
|---|---|
| videocortex | Drop a clip. See which cortical regions TRIBE v2's average subject would fire. Encoding, not decoding — local instrument around Meta TRIBE v2 (Metal, preflight, plates) |
| videocortex-spark | Same instrument on NVIDIA DGX Spark (GB10 / CUDA 13). UMA from meminfo, CIRC-fixed runs view, plus export / sonify
|
Category page: miscellaneous-research
Long-form companion writing to the repos above — 34 articles spanning 2023-04 to 2026-07, covering AI security governance, agentic pentesting, OSCAL-as-code, custom silicon economics, and zero-trust AI coding.
See timeline for the chronological list, or Index for all article pages grouped by category. Articles are also linked from each category page under the "Related Articles" heading.
Across all 45 live repositories, several architectural patterns recur
(counts from the stacks: frontmatter of repo stubs, see Sitemap-Stacks):
- Agentic — multi-agent orchestration / tool-using agents (24 repos)
- LangGraph / LangChain — dominant orchestration framework (24 repos)
- Apple MLX — on-device inference on Apple Silicon (16 repos)
- CLI / Tooling — command-line utilities and workflow glue (13 repos)
- Compliance — regulatory frameworks, controls mapping (12 repos)
- OSCAL — NIST OSCAL data model (SSPs, profiles, controls) (9 repos)
- Pentest — offensive security, red-teaming, vulnerability discovery (8 repos)
- NVIDIA DGX Spark — CUDA / local vLLM ports of the desk fleet (8 repos)
- RAG — retrieval-augmented generation, vector stores (5 repos)
- Converter — file-format conversion, OCR, doc-to-markdown (3 repos)
- MCP — Model Context Protocol servers / tooling (2 repos)
- Pydantic for data validation (nearly universal)
- Typer + Rich or Click + Rich for CLI interfaces
- Karpathy LLM Wiki pattern for knowledge bases (this wiki)
- Dual-platform siblings — Apple Silicon and DGX Spark trees of the same product: driftlab, screen-lens (mlx/oMLX vs vLLM), videocortex (Metal vs CUDA). TSLIT collapsed to one live spine (tslit); there is no public mlx TSLIT-DSPy repo.
Wiki structure: see SCHEMA | Full index: Index | Visual map: Sitemap | Change log: Log | Stacks view: Sitemap-Stacks