KoishiAI Lab
Work we measured or tested ourselves, not summaries of other outlets’ news. Each piece states the method it used; some come with raw data you can download and check.
- New
How many more tokens does Thai take than English?
Measured on 1,997 news sentences translated into Thai with three real tokenizers, plus our own articles and the raw data.
Read → -
Four Thai baht-text libraries tested on 3,289 amounts against LibreOffice
We compared the latest thai-baht-text, bahttext (npm), PyThaiNLP and bahttext (PyPI) with LibreOffice's BAHTTEXT. Ordinary amounts read almost identically, but every library has cases to watch: zero, negatives, one satang and three decimal places.
Read → -
Thailand's Royal Gazette set a record in 2025, and nearly 6 in 10 notices are bankruptcy cases
We counted every Royal Gazette notice in the Open Law Data Thailand dataset: 49,127 in 2025, the most since 1885; about 4 in 10 land on Wednesdays; nearly 6 in 10 are bankruptcy notices; and in early 2025 a backlog of over 3,000 local ordinances was published at once. Plus the data limits to know before you use it.
Read → -
Measuring Thai Text Programmatically: Why .split() and crfcut Lie to You — and How the "Wak" Saves the Day
Lessons from building a naturalness meter for AI-written Thai articles: the silently-broken word and sentence tools, and how to measure Thai rhythm in the unit the language actually marks.
Read → -
Local LLM Benchmark on a 48 GB Dual-GPU Rig: What Actually Runs in 2026
We ran Qwen3 27B, 32B, 35B-A3B, and 80B on an RTX 5090 + 5080 box to find the real sweet spot for local AI in 2026. Here is which model fits which job — and which one to drop.
Read → -
Fine-tuning Thai OCR for production: seven full fine-tunes that barely helped, a LoRA that works in practice, the guards built with it, and a leaky test set
A real record of fine-tuning a 2B Thai OCR model to read business documents: seven rounds of full fine-tuning, the general ability it lost, the switch to a LoRA that made the live system work, the loop guards developed alongside it, and a train/test leak we only just found, with the numbers recomputed.
Read →