Language datasets
Corpus building from books and archives — segmentation, labeling, licensing, provenance.
World hub · data & technology
Connecting local languages and heritage to the world.
A data engine for endangered languages and manuscript digitization — built in Sumatra, for the world.
Services
Five concrete capabilities, one pipeline — from raw archives on a shelf to machine-readable language data, and the infrastructure to serve it.
Corpus building from books and archives — segmentation, labeling, licensing, provenance.
Photo-to-searchable-text pipelines for regional-language documents.
Text and speech translation built on top of the Li Niha dataset.
Page-by-page restoration and print preparation of heritage scriptures.
Self-hosted inference, OCR API, and encrypted tunnels — private by default.
Project 01
Nias (Li Niha) is spoken by ±1 million people on Nias Island, North Sumatra — with near-zero NLP support. This dataset changes that: a clean, licensed, provenance-tracked foundation for teaching machines to read and translate a language the industry forgot.
CRNN + CTC baseline, trained end-to-end on a consumer CPU — no GPU.
30 source books from the “Li Nono Niha” series, photographed page by page.
41,971 labeled word images across a 59-character alphabet, incl. õ ö ô ’.
CRNN + CTC baseline trained on a consumer CPU — 98.8% character, 97.7% word accuracy.
71,339-word corpus, 7,807 unique words — every example records its book and page.
For tourism and surfing destinations across Nias — a bridge between visitors and the local community: text and speech in, text and speech out.
Turn Nias teaching books, church and government archives into searchable digital text — from shelf photo to corpus.
Project 02
Buddha Names Sūtra · restoration & print preparation
Restoration and print preparation of the Buddhist scripture 佛說佛名經, Volume I & II — 28 Katern per volume, CorelDRAW sources, 380 × 260 mm print sheets.
Red Hanzi in vertical columns, blue pinyin at their side — the visual grammar of the scripture, preserved exactly.
Every file is SHA-256-hashed before and after work — sources are never mutated.
Every job happens in its own separate, timestamped run.
Every page is saved, closed, and reopened before it is counted as done.
Every page is personally approved by the owner before the next one begins.
100% Hanzi accuracy
FAIL — characters got duplicated
Key finding: reading the vector text directly from the source file is 100% accurate — outperforming every raster OCR engine.
Infrastructure
Self-hosted GPU inference node (48 GB VRAM) running local OCR & LLM models.
OCR API (PaddleOCR PP-OCRv6) on a Linux server — localhost only, reached through an encrypted tunnel, never publicly exposed.
HunyuanOCR handles the Project-15 restoration line — 100% Hanzi accuracy, best of four OCR engines benchmarked.
An encrypted tunnel bridges the private network to the internet — irwan.world DNS is the public doorway.
Our Network
irwan.world is part of a four-site ecosystem — data & technology, print, network infrastructure, and AI.
Trusted print shop in Medan — invitations, banners, business cards, books, and other print needs.
Visit →Network infrastructure, e-voting, cloud ECS/EDS, plus web design, hosting, and domains.
Visit →AI studio — AI training, model fine-tuning, and web/desktop/mobile app development.
Visit →Contact