Ilnar Salimzianov's Personal Site

Work with me

NLP / ML engineer specializing in speech-to-text and low-resource languages. I take language and speech models from notebook to production — fine-tuning, ONNX optimization, APIs, deployment, and load-testing — and I do it for languages the big cloud providers handle poorly or not at all.

Computational linguist by training (M.Sc., University of Stuttgart, 2017), with a decade building NLP, speech-to-text, and MLOps systems across industry and research.

Availability

About 20 hours per week, across 3 days — a deliberate, stable commitment, so the rest of my week can go to my own language-technology work at taruen.com and selimcan.org. You get a focused senior contractor who is in this for the long run, not someone who will churn the moment a full-time role appears.

Remote, worldwide. My working hours overlap European business hours and US-East mornings. I'm based in Istanbul and relocating to Serbia — putting me squarely in the European working day.

How I can help

Speech-to-text for your language or domain

Fine-tune, evaluate, and deploy automatic speech recognition (Whisper and others), including languages not supported out of the box. Honest WER/CER benchmarking, not vendor hand-waving.

ML models from notebook to production

Take a working model and make it deployable and affordable: ONNX conversion and quantization, a FastAPI service, Docker, load-testing, and cost optimization. (At ProWritingAid, ONNX conversion and quantization let a workload the team had scoped for 8 GPU machines run on 1 — plus a warm standby for redundancy — at negligible accuracy loss.)

NLP data pipelines at scale

Scrape, clean, and structure large text corpora for NLP/LLM training and search — for example, an end-to-end ingestion pipeline over 20M+ USPTO trademark records feeding a multi-layered Elasticsearch similarity engine.

Low-resource & multilingual NLP

Morphological analysers, machine translation, data collection, and evaluation for Turkic, Caucasian, and other under-resourced languages. Eight publications and several shipped open-source tools in this area.

Selected results

• Shipped ProWritingAid's premium-tier tone-detection feature — training, fine-tuning, Azure ML deployment, and ONNX optimization.
• Built an end-to-end pipeline over 20M+ USPTO trademark records with phonetic, orthographic, and cross-lingual similarity search.
• Fine-tuned Coqui STT for Kazakh (arXiv:2107.10637) and Whisper for Laz, and launched Common Voice for Kazakh and Tatar.

Get in touch

The fastest way to start is a short message describing what you're building and the outcome you need.

Home | Work with me | Résumé | Projects | Publications | Talks | Now | Email | Reading log | Movies log