CTO & Principal AI Architect.
Anthropic Claude Certified Architect - Foundations · November 2026.
I build open-source tools for AI-agent reliability — small, tested, documented, with honest limits.
| Repo | What it does | Try it |
|---|---|---|
| agenthub | See every MCP server, skill and instruction file your coding agents actually load (21 supported) — and what's broken. Read-only, no account. Plus a sync control plane with drift detection. | npx @shmindmaster/agenthub doctor |
| heatlot | AI reads supplier certificates; deterministic rules decide; cited, signed receipts. | npx heatlot demo |
| subtext-scan | Find hidden prompt-injection text in PDF/DOCX/HTML before your agent follows it. The MCP server withholds the payload. | npx subtext-scan scan file.pdf |
| replyonce | Stop your agent replying, refunding or ticketing twice. Fencing tokens + effect ledger. Node + Python, zero runtime deps. | npx replyonce demo |
| fakeplane | Stateful fake Zendesk/Slack/Linear/HTTP backends for agent evals that block bad releases. | npx fakeplane demo |
| safeguards | Dataset: 356 publicly collected AI system prompts scored against 23 written safeguards. None states even half. Data CC BY 4.0. | Hugging Face |
They form an agent-reliability stack: screen untrusted documents (subtext-scan), make side effects at-most-once (replyonce), test against stateful fakes before release (fakeplane), and keep every coding agent's config in sync (agenthub). heatlot is the worked example; safeguards is the evidence that prompts alone don't enforce controls.
Writing: saroshhussain.com.



