← Feed

Posts

• View post

  Some of the unique characteristics of developing an OCR recogniser for Armenian. https://tetrak.dev/articles/2026/09/shipping-the-armenian-recogniser/ Now public. Model weights and training data on Hugging Face. Code on GitHub. Permissive licensing: MIT and Apache 2 for code and weights. The training data corpus inherits Creative Commons share-alike licensing due to its sources. Scoring pretty well already!

• View post

  Come on a tour of our open source Armenian OCR projects, how to help make it even better, and how to copy it and train an OCR recogniser for your own language. Or just get in touch and maybe we can do it for you! https://tetrak.dev/articles/2026/09/built-to-be-forked/ #armenian #ocr #ai

• View post

  Even Fable 5.1 is quietly slipping “quietly” into everything. I wonder whether things like that and “load-bearing” are the AI equivalent of EVs having speakers to make unnecessary noises so that humans feel less threatened around them.

• View post

  @thomasfuchs In the cloud migration space, I’ve had that conversation. “Oh you mean replication?” … “No, no, no this is a hydration project!” … proceeds to describe replication.

• View post

  I just completed some useful analysis of new vision-language models for OCR using PaddleOCR-VL and GLM-OCR. Got great results on difficult images with Paddle, but the article explains how GLM-OCR was guilty of massive fabrication! https://tetrak.dev/articles/2026/09/the-engine-that-made-it-up/

• View post

  Today, Claude Code started throwing snippets of Armenian into its responses to me!

• View post

  Should your backups live in a separate cloud subscription? I looked into what real attacks on cloud backups tell us about keeping the recovery copy in the same Azure subscription as production — where immutability holds, where a subscription boundary helps, where only a separate tenant does, and what the regulators and frameworks actually ask for. https://scattercode.dev/2026/10/should-your-backups-live-in-a-separate-cloud-subscription/

• View post

  A "turned sort" - a new term I learned just now. The second letter կ was accidentally placed upside down by the printer. This is an 1873 issue of Մշակ magazine, which used to be printed in Tiflis (now Tbilisi, Georgia). I'm currently transcribing a few pages of that magazine to train the Tetrak Armenian OCR on a font/typesetting variant it hasn't yet been trained with. Right now, its abilities on older texts are not as strong as 20th century, so I'm exp...

• View post

  I have been doing some work that relies on determining whether text is written following Classical Armenian orthography or the reformed variant more commonly used in the Republic of Armenia. Interestingly, it mostly came down to counting the letter ւ. https://tetrak.dev/articles/2026/10/classical-or-reformed-armenian-by-one-letter/

• View post

  Is the name of that great Armenian writer of poetry and fairy tales Tumanyan? Tumanean? Toumanian? The term “fold” is all over the Tetrak codebases and I’ve used it a couple of times in articles, so I thought I’d write about it. Although I should warn that as it developed, this article pivoted into transliteration and Romanisation with something of a potted history of Armenian orthography reforms and what the diaspora thought of them. I did enjoy diving into this rabbit hole, so hopefully you mi...

• View post

  Today, I found that the ISO standards for transliterating Georgian and Armenian (and I'd guess more) changed earlier this year. So I paid for the new 2026 editions and implemented them in tetrak-translit. The Georgian standard has some major changes (and a bug/defect). https://tetrak.dev/articles/2026/10/what-changed-in-iso-9984-and-iso-9985/

• View post

  When libraries around the world catalogue Armenian content, they usually use transliteration to provide a phonetic version in the Latin alphabet (aka Romanisation). tetrak-translit is a new Python library for Eastern and Western Armenian using LoC/ISO/etc standards. It can also normalise them all to folds for search. We added Georgian and if you’d like to see more alphabets/scripts supported, let us know. https://github.com/scattercode/tetrak-translit

• View post

  A fun probability rabbit hole that I didn’t imagine I would need to explore until last week. How can a computer tell whether a page of Armenian text is sideways or upside down? https://tetrak.dev/articles/2026/10/which-way-up-armenian/

• View post

  Romanisation / transliteration of Syriac script is an interesting challenge for library catalogues. We have a decent first version of a Syriac search fold that is giving decent results. https://tetrak.dev/articles/2026/10/searching-syriac/

• View post

  Be careful if you have automated OCR pipelines using Marker processing images you can’t be sure about. That and its dependency, surya-ocr are currently locked to Pillow <11 which has 3 critical, 8 high, and 4 moderate vulnerabilities. I’ve pulled it as an option from tetrak until I see that it’s picking up Pillow >=13.