About Nalin

What the models read

Founded in Seoul in February 2022

A Korean commerce app's head of engineering owns its email templates, and since October 2026 a coupon notice sent without consent can cost a share of revenue. The Network Act's surcharge reaches 6% of revenue related to a violation, and KISA's March 2026 guide counts unrequested coupon and points notices as advertising.

Nalin is an email API for Korean apps, run from Seoul, whose own models read every template before a named person approves it. A classifier proposes informational or advertising, a drafting model writes templates in Korean and English from an event's fields, and a forecaster predicts inbox or spam at Naver Mail, Daum and Gmail. Advertising mail carries '(광고)' and a bilingual opt-out and reaches only recipients with a dated consent.

Whether a template is informational or advertising is a call rules settle only at the edges, and Nalin's classifier works the middle. A password reset or a segment broadcast settles by rule; a points notice turns on who asked and where the points came from. The classifier reads the subject, the body in both languages and the event's fields, and returns a class with the sentences that decided it.

No new or changed template version leaves Nalin until the person named as its owner approves it, and no API key can approve. Below a 90 percent confidence line, or where a template mixes both classes, the classifier proposes no class and the owner decides. From customers who opt in, each decision becomes a labeled pair Nalin keeps: the deciding passage beside the class, merge values removed.

Two moments fill Nalin's GPUs: a new customer's template library read at once, and every customer's templates re-read when a regulator's guide changes. An app sending 2 million emails a month brings about 140 templates, classed within hours. A model API abroad would be one more overseas transfer to disclose, so the classifier and drafting model run on cloud GPU capacity we control in Seoul, beside mail servers on cloud infrastructure in Seoul.

Nalin's classifier and drafting model are open-weight models it fine-tunes, today on GPU capacity rented hourly in Seoul. The classifier learns from a corpus Nalin's privacy team wrote and labeled from KISA's guide and the Network Act decree's Annex 6, and from opted-in owners' pairs. In April 2027 the hourly rental becomes a twelve-month GPU commitment in Seoul, with quarterly retraining on those pairs; the small forecaster stays on ordinary servers.

Nalin Co., Ltd. was registered in Seoul in February 2022 to send Korean apps' transactional email from servers in Korea. Delivery at Naver Mail and Daum filled its first two years; the forecaster followed Naver's July 2024 sender rules, and the classifier KISA's March 2026 guide.