use case · multi-model · DistilBERT + MarianMT
Before you send a support reply to a customer in another language, two things matter: that it reads the way you intend, and that it's in their language. This pipeline does both, on-device: DistilBERT checks the tone of your English draft, then MarianMT localises it. Catch a reply that reads harsher than you meant before it ships — and nothing leaves the tab.
Each model initialises on its own below (cached copies auto-start, no button).
DistilBERT — sentiment (~65 MB)
MarianMT — translation (~105 MB)
step 1 — you
step 2 — DistilBERT
step 3 — MarianMT
Sentiment analysis and translation are each useful alone. Chained, they're a quality gate for multilingual support: the tone check flags a reply that reads negative before you commit to sending it to a customer who may already be frustrated — and the localization means they read it in their own language. No server, no API bill, and the customer's message never leaves your device.