Konstantin Kolomeitsev
AI/LLM Architect and Applied AI Researcher
I help companies design and implement LLMs, AI agents, RAG, local models, and business-process automation: from the first audit and strategy to a working system.
- Based in
- Almaty, Kazakhstan
- Current role
- CAIO @ Sergek Development
- Founder
- Loqira Labs
- AI implementation and AI audits for companies
- CAIO-level advisory and AI strategy
- Corporate AI training (delivered in Russian)
- Research collaborations
A published study on transferring knowledge from a large model to a small one via Enhanced Cross-Attention.
Read on arXivA released app for Windows and macOS: speech recognition runs locally on a CPU, free.
Open LevitanAn original architecture that beats Nvidia's Cosmos Tokenizer on UVG by average Y-PSNR.
See the workHow I can help
Four ways to work with me: implementation projects, consulting, team training, and open research.
Process audit, task selection with measurable impact, a pilot, then production: LLMs, RAG, AI agents, local models.
A working session on your task: where AI applies, what the architecture looks like, what it costs, and what to do next.
Programs for executives, developers, and staff. Delivered in Russian.
Published work, models, and open code: LLM Modules (arXiv), MPT-VC, Qwen3.5 GEC.
Architecture plus delivery, engineered for production
What I build and how I de-risk it.
Multi-agent pipelines, retrieval strategies, evaluation harnesses, guardrails, and privacy-by-design.
Process mapping, an automation backlog, KPIs and ROI: shorter cycle time, lower OPEX, better conversion.
Regression testing for prompts and tools, observability, and cost/performance tuning at scale.
Risk management, privacy, monitoring, bias mitigation, and clear accountability for production AI.
Finetuning, quantization, and CPU inference: capable models that run on your hardware, without a cloud.
AI strategy audits, technical and product due diligence, integrations, and executive alignment.
Published work, not slides
Papers, open weights, and shipped products that back every claim on this site.
Knowledge transfer from a large model to a small one, no distillation, trained for under $10.
A neural video codec on original operators; beats Nvidia's Cosmos Tokenizer on UVG by average Y-PSNR.
A typo, spelling, and punctuation corrector for Kazakh, Russian, and English that runs on a CPU.
On-device voice input for Windows and macOS: local, private, free.
Frequent questions
Who is Konstantin Kolomeitsev?
A Chief AI Officer (CAIO), AI/LLM architect, and applied AI researcher from Almaty, Kazakhstan. Founder of Loqira Labs, author of the LLM Modules study (arXiv:2502.08213) and shipped products, including the on-device voice input Levitan.
How can you help my business?
I assess where AI applies to your processes, then design and implement the systems: LLM assistants, RAG over company knowledge, AI agents, local models. I also run AI audits, consulting sessions, and corporate training.
Do you work with companies in Kazakhstan?
Yes. I live in Almaty and work with Kazakhstani companies on-site and online. Companies from other countries: online.
Do you run AI training?
Yes, corporate programs for executives, developers, and staff. Training is delivered in Russian.
Where can I see your research?
The Research page collects the work with papers, weights, and code: LLM Modules, MPT-VC, Context Merging, Qwen3.5 GEC. Profiles: arXiv, Hugging Face, GitHub.
How do I contact you?
Email libr@bk.ru is the fastest channel. LinkedIn works too. Cooperation formats are listed on the contact page.
Discuss your task
Write a couple of paragraphs about the process you want to automate. I will reply with what I can do and where to start.