Hi from London. I build the guardrails AI agents ship without.

Lounge found me through my GitHub, so that's where I'll start.
A while back I was adding an AI assistant to a customer support tool. Routine work. I typed a slightly odd question and the model dumped part of its own system prompt into the chat.

I hadn't done anything clever. Which meant anyone could do it on purpose.
That afternoon turned into prompt-protection. It started as a prompt-injection filter. Version 3 is a tool-call guard for JavaScript agents. It tracks where every piece of data came from, and untrusted data never reaches a network, email, exec, file or payment call. It's listed in the official MCP registry.

Four smaller libraries sit around it, all on npm:
🛡️ mcp-policy-guard: decides which MCP tool calls are allowed, which need a human, and what a response may contain
🌊 streamguard-react: error boundaries for streaming LLM UIs that keep the partial tokens when a stream drops
📏 ragbench-lite: golden-question evals for RAG pipelines that fail CI when retrieval regresses
⚛️ react-crash-guard: React error boundaries that catch the async errors a normal boundary misses
Day to day I'm a senior full stack AI engineer. Five years of putting LLM features in front of real users, MSc in Software Engineering from Kingston. More at mozia.dev.

I'm here for founders shipping agents, and for the London meetups.
If your agent can send email, I'd like to hear how you stop it sending the wrong one.