Chaukidar
Vercel
Unicorn Track
Chaukidar helps teams evaluate whether AI models remain safe beyond English-only testing. It audits LLM responses across South Asian languages and harm categories such as fraud, cyber abuse, self-harm, hate, and violence. The platform compares English seed prompts, translation-baseline prompts, and native-adapted prompts to reveal where model safety behavior changes across language, culture, and phrasing. Built for AMD Hackathon Track 3, Chaukidar uses AMD ROCm/vLLM notebook results for compute-backed audits, supports Fireworks model comparisons, re-judges imported outputs with a multilingual judge, and generates dashboards, risk heatmaps, track comparisons, and audit reports.
Category tags:Explore more applications
Upcoming AI Hackathons
For Innovators & Creators
For Innovators & Creators
Explore more applications