Share ideas. Build tools for practice.

SIDE Framework

Four questions to uncover misalignments between responsibility, authority, roles, and decision-making.

A lens for retrospectives, Agile, and digital transformation: distinguish structural problems from incidental friction.

CORE Framework

A design compass using Context, Object, Reaction, and Evolution to examine architecture and strategy.

Connects purpose, responsibility, consequences, and future change with the practical perspectives of Seven Ark.

Inverse Value Canvas

Reframes missing features and omitted elements as opportunities for value, differentiation, and collaboration.

Moves through Missing, Harmed, Inverted Ally, and Strategic Fit to identify who benefits from an absence and why.

SecuLens

A security assessment CLI distributed via npm, PyPI, and Packagist. Matches SBOMs against vulnerability data, checks license policy, and runs static analysis — all with cited evidence.

Matches SPDX/CycloneDX SBOMs against the OSV advisory database, parses complex license expressions (AND/OR/WITH), and generates localized Word reports for clients. Distributed with the same design across three ecosystems: TypeScript, Python, and PHP.

Web Security Hub

A local tool combining web-security assessment and form regression tests, with Playwright exploration, ZAP integration, and static analysis.

Supports browser scenarios, form test cases, and findings with evidence and remediation guidance.

Codes Doctor

A VS Code extension that diagnoses god classes, technical debt, security issues, and dependency structure directly in the workspace. No external server — the analysis engine runs entirely in-process.

Supports 11 languages and traces external input to dangerous operations within functions. Shareable HTML reports support light and dark modes.

SBOM Vulnerability Scanner

A VS Code extension that generates a software bill of materials (SBOM) and checks it for known vulnerabilities, right in the editor. Results export as HTML, JSON, or Word.

Auto-installs Syft and Trivy, detects multiple npm projects in a monorepo, and produces a consolidated report — with Japanese-language output supported.

LLM Security Scanner

LLMScan analyzes Python, TypeScript, JavaScript, and PHP locally, tracing external input and LLM output to risky SQL, shell, and HTML operations.

Includes live diagnostics, path details, Quick Fix, custom rules, and baselines. The CLI supports changed-file scanning and SARIF export.

Model Audit Lite

A lightweight CLI that checks a Hugging Face model's safety in about five minutes, with zero infrastructure — flagging pickle weights, custom code, and other arbitrary-code-execution risks.

Combines three independent checks: file-distribution auditing, before/after conversion-consistency comparison, and a known-pattern safety probe suite. Not a replacement for heavyweight tools like Meta's CyberSecEval — scoped narrowly for whoever is doing the conversion to self-check their own work.

AgentDojo (Japanese)

A Japanese adaptation of AgentDojo, the prompt-injection benchmark for LLM agents — rebuilt for Japanese context, not just translated.

Provides the same four suites (banking, Slack, travel, workspace) with 97 user tasks and 31 injections, including 25 Japanese-specific attack patterns such as keigo-based and fake system-notice injections.

Ruri Atlas

A browser-only Japanese/English semantic search tool. Computes embeddings via WebGPU and plots semantically related sentences on a 3D map.

Supports switching between embedding models including ruri-v3, with a built-in feature to visualize how Japanese prompt-injection patterns distort the embedding space. All processing stays local — nothing is sent externally. Published as a Hugging Face Space — try it directly in the browser.

Agent Defense Radar

A visualization tool for understanding and defending against indirect prompt injection. Shows real attention patterns from a live model and lets you test your own endpoint.

Pairs a recorded Model X-Ray (attention and response, guarded vs. unguarded) with a Live Model Test against your own OpenAI-compatible endpoint. Published as a Hugging Face Space — try it directly in the browser.

2-bit quantization that fits large models on a 16GB Mac

Project Pyrite

Project Pyrite is a sensitivity-aware ultra-low-bit quantization pipeline that reallocates bits to the most sensitive layers. It brings Qwen3.8-27B, llm-jp-4-32B-A3B (MoE), DeepSeek-Coder-V2-Lite and others down to 2 bits, cutting KL divergence from the original by up to ~22% at exactly the same byte budget as standard 2-bit quantization, so they run on a 16GB Apple Silicon Mac.

GGUF builds for llama.cpp, Ollama, and LM Studio accompany the MLX builds. Also published: the MoE-Stream engine (pyrite-stream), which streams only the needed experts from NVMe, and a streaming container format for FLOW Studio. The paper is available on Zenodo.

Japanese text embeddings

Ruri v3

The Japanese text-embedding model ruri-v3 — the foundation for on-device semantic search, similarity search, and RAG — converted across four runtimes (Core ML, MLX, Core AI, TFLite) and four sizes (30M/70M/130M/310M), so it runs server-free on iOS, macOS, or Android.

A 310M reranker variant is also available in TFLite, MLX, and Core AI. These on-device ports pair with the WebGPU conversion used in the browser-based Ruri Atlas.

Japanese search reranking

Japanese Reranker v2

The Japanese reranker japanese-reranker-v2, which reorders search results by semantic relevance, converted in two sizes (xsmall/small) across five runtimes (Core ML, MLX, Core AI, TFLite, ONNX for web), so the same model runs from the browser down to mobile.

Paired with ruri-v3 embeddings, this completes a full server-free Japanese RAG pipeline entirely on-device.

Run a Japanese reasoning model on a Mac

LLM-jp 4.1 8B Thinking

The reasoning-focused model llm-jp-4.1-8b-thinking from the LLM-jp project, distributed as a 4-bit quantized MLX build and a GGUF build — run Japanese chain-of-thought responses on Apple Silicon without a server.

The previous-generation llm-jp-4-8b-thinking is also available as a 4-bit MLX build, making comparison and migration easier.

A coding-focused model that runs on a Mac

Qwen3.5 4B Agentic Coder

A Qwen3.5 4B model specialized for agentic coding assistance, 4-bit quantized and optimized for Apple Silicon via MLX. The most-downloaded model in this collection.

A lightweight, practical option for trying agentic code generation and review directly on a Mac, with no server round-trip.