PsiGuard – real-time hallucination monitoring for LLM apps

Hey HN,

I’ve been building LLM apps for a while, and one thing kept bothering me is that models will say completely wrong things with total confidence. Not edgy wrong or funny wrong, just confidently fabricated citations, broken reasoning, medical summaries that look plausible but aren’t. The kind of stuff that’s fine in a chatbot but sketchy in production.

So I built PsiGuard. It’s a lightweight SDK that wraps around your existing LLM call and scores the output for likely hallucination patterns. It returns structured metadata (confidence score, flags, anomaly signals) and lets you decide what to do — log it, warn, or block.

The goal isn’t to replace the model. It’s to add a safety layer.

Basic usage looks like:

from psiguard import monitor response = monitor(llm_output) print(response.score, response.flags)

It’s early, but it’s working well in my testing across general Q&A and domain-specific prompts.

Repo: https://bumpr-ai-safety-rafj.onrender.com/ SDK: pip install psiguard

I would very much appreciate feedback, especially from anyone running LLMs in production or dealing with hallucination mitigation in real apps.