I build AI that must survive production: affordable, trustworthy, accountable. My work is organized around one belief — AI entering enterprise production has to pass three gates at once, and I build a system for each gate.
- aiitg — AI Input Trust Gateway: hidden-content auditor for documents fed to LLMs and agents (zero-width chars, white text, hidden sheets). 102 tests.
- keysmith — Go MCP server + CLI for agent-safe secrets: age-encrypted, masked views, self-healing rotation. Plaintext never enters agent context.
- fusion-cache — LLM API caching middleware: exact-match → semantic → upstream prefix-cache. 91 tests, up to 97% cost cut.
- dsh-llm-cache — LLM response cache merged upstream into DeepSeek Harness (
@deepseek-ai/dsh-llm-cache).
- takumi — model-agnostic platform for verifiable agentic software delivery: Manager / Executor / Auditor, verify-fail → repair → re-verify. 87/87 tests; 23-task eval 96% first-pass → 100% final.
- proctor — agent eval: score, audit and explain agent trajectories, with metrics and causal failure attribution.
- decision-workbench — expand the decision space before you decide: multi-model committee + human final call. Hidden-gold eval: 26% → 78% → 100% (12/12).
- decision-space-expansion — the methodology behind decision-workbench, as a portable Agent Skill (EN/JA). Works with Claude Code, Cursor, Codex and Gemini CLI.
- agency-agents-ja — 404 production-ready AI agents for Japanese dev teams (125 Japan-original), for Claude Code / Cursor / Copilot.
- skills-ja-JP · superpowers-ja — Japanese localizations of engineering skill packs for AI coding agents.
www.sscodeai.com — portfolio: story, résumé, project deep-dives
sscodeai.com — product docs, tutorials & Japanese-market agent toolkits