PromptRails

Local-first, honestly-benchmarked AI for privacy.

We build small models that run on your own machine — no cloud, no API, no data leaving the device — and we test them on independent, adversarial benchmarks so the numbers actually mean something.

Shipped

🔒 piimask-qwen2.5-0.5b — a 0.5B model that detects PII in English and Turkish, fully on-device. Turkish PII (names, addresses, TCKN, IBAN) is underserved by English-first tools; piimask is built for it.

📊 piimask-eval-tr-en — an independent, adversarial bilingual PII benchmark. We publish the set we test on — including the cases we still miss.

What we care about

• Local-first — privacy by default; nothing leaves your machine.
• Honest benchmarks — independent sets, disclosed limitations, reproducible numbers.
• Open — Apache-2.0 weights, open eval data.

🌐 promptrails.ai · 💼 LinkedIn