Crash Test
About this project
In one sentence, what are you building?
A self healing test harness that attacks your voice agent with a swarm of AI callers, grades every call, rewrites the prompt to fix the failures, and attacks again.
What problem are you solving?
Teams ship voice agents after a few manual test calls. Nobody checks what happens when a caller lies about who they are, injects instructions, or asks for advice the agent must never give. The failures show up in production, where a leaked account detail or a wrong answer costs customer trust and creates compliance risk. Manual testing takes hours and still misses these cases. Crash Test runs ten adversarial callers in parallel in about a minute, shows exactly which attacks worked, and repairs the agent on the spot.