Steve Swain

Certifai - Catch AI hallucinations before they reach your users

by
Most AI hallucination tools check if content is true. But the failures that actually break agent pipelines are often structural: invented JSON fields, mismatched tool outputs, or prompt injections hidden in retrieved content. This suite runs four suppressors in one API call: grounding checks, injection detection, JSON schema validation, and tool‑response verification. Available as a REST API and MCP server. Free tier (500 requests/month) at certifai.dev

Add a comment

Replies

Best
Steve Swain
Maker
📌
I built this because I kept running into the same failure mode when working with AI agents: not truth hallucinations, but structural ones. The model would invent a field in a JSON response, claim a tool returned something it didn’t, or cite a source that wasn’t in the retrieved set. Sometimes a prompt injection was buried inside retrieved content and slipped through because it looked “trusted.” By the time you notice, the damage is already downstream. I wanted a single checkpoint that runs before agent output reaches users, something that enforces grounding, validates schemas, checks tool‑response integrity, and catches injection attempts inside retrieved content. That became the four suppressors in this suite. Claude Code made it fast to iterate: I used it to generate validators, test edge cases, and refine the API contract until the suppressors behaved deterministically. If you’re building agents that rely on tools, retrieval, or structured outputs, I hope this helps you avoid the same failures I kept hitting. Happy to answer questions about how any of the suppressors work or how to integrate them into your pipeline.