Darmm AI

Darmm is an independent R&D effort focused on language and document AI for underserved languages — starting with Kazakh and the broader Central Asian context. Maintained by R3iwan.

The focus is the neglected part of the AI landscape: languages and domains where global open-source models quietly drop in quality — where tokenizers fragment text, models drift into a dominant neighbor language, and local script, context, and terminology are treated as edge cases.

Website: darmm.kz — currently AI tutoring, gradually becoming a project showcase.

What I'm working on

• Text / NLP — LLM fine-tuning, embeddings, classification, and generation for low-resource languages, with a focus on legal and technical terminology. Current work centers on Kazakh and Russian.
• Tokenization & language adaptation — measuring and closing the gap between how global models handle underserved languages versus how they should: tokenizer efficiency, vocabulary extension, continued pretraining on native-language data. A methodology built on Kazakh, designed to transfer.
• OCR & document AI — recognition of printed and handwritten text in underrepresented scripts, adapted for legal, educational, and government documents; extraction and validation pipelines on top of it.
• Retrieval & RAG — embedding models, hybrid search, and RAG/GraphRAG pipelines tuned for non-English corpora, evaluated against curated golden sets.
• Benchmarks & datasets — open datasets and honest, reproducible evaluation reports of where current models actually stand on languages the leaderboards ignore.

Most work is published as open models, datasets, and benchmarks on Hugging Face. Production-specific components stay closed.

Technical focus

• LLM adaptation: LoRA/QLoRA fine-tuning, vocabulary extension, domain-specific embeddings, instruction tuning for low-resource settings.
• OCR: classical and transformer-based OCR pipelines, support for multiple scripts and alphabet variants (currently Kazakh Cyrillic and Latin), document layout and field extraction.
• Retrieval: embedding benchmarking, hybrid scoring (BM25 + dense), rerankers, RAG and GraphRAG pipelines.
• Inference: vLLM serving, quantization (AWQ, GGUF), latency optimization for production deployment.
• Evaluation: CER for OCR, retrieval and generation metrics for RAG (including LLM-as-judge calibrated against human labels), honest reporting of where models fail.

Why start with Kazakh

Kazakh is the native language of over 20 million people, the state language of Kazakhstan, and the language of a growing digital economy. In AI research, it is consistently treated as an afterthought — and it is far from alone.

Tokenizers of major open models fragment Kazakh text far more than English or Russian, degrading both quality and cost. General-purpose LLMs either slip into Russian when prompted in Kazakh, or produce degraded output. OCR systems regularly confuse Kazakh Cyrillic with Russian. Domain terminology — legal, governmental, technical — is where the drop is sharpest.

The same failure pattern repeats across dozens of languages worldwide: a dominant neighbor language the model falls back to, a script the tokenizer wasn't built for, domain vocabulary nobody benchmarked. Kazakh is where Darmm builds and validates the playbook — systematic benchmarking, open models, honest reporting — with the explicit goal of making the methods, tooling, and evaluation practices reusable for other underserved languages.

This is not a niche academic gap. It's a practical failure affecting real users and real products, in languages that deserve better infrastructure.

Coming later

• Kazakh ↔ Russian / Kazakh ↔ English translation, with the pipeline generalized to other language pairs
• Multimodal document understanding (vision + language)
• Domain-specific verticals — legal, medical, education
• Extension of the adaptation and benchmarking stack to other Central Asian and low-resource languages
• Speech (ASR/TTS) — deliberately parked until the text and document stack is solid

Status

Darmm is an early-stage R&D effort, not a finished product line. Models and writeups are published as they're ready, with honest baselines and known limitations. Reach out via github.com/R3iwan.