README.md

metadata

title: README
emoji: 🌍
colorFrom: red
colorTo: gray
sdk: static
pinned: false

Craftloop AI

Open-source, CPU-friendly small models for Indian-language AI safety.

Most open guardrail models are English-only, and they often block normal Hinglish users or miss Hinglish attacks. Craftloop AI builds small, practical models and datasets for English, Hinglish and Hindi.

Models

• hinglish-prompt-guard: detects prompt injection & jailbreaks in English, Hinglish and Hindi. Hinglish F1 0.97 · false positives 0% (vs 31.5% for an English-only guard)

Datasets

• hinglish-prompt-injection: synthetic Hinglish/Hindi prompt-injection data with a template-disjoint test split

Coming next

• Indian PII redactor (Aadhaar, PAN, UPI, phone numbers) for safe LLM use
• An LLM safety gateway: guard → redact → LLM → restore

All models and data are open (MIT). Feedback and contributions welcome.