Skip to content
Konstantin KolomeitsevKonstantin Kolomeitsev
Discuss a project

Projects, products, and open source

Products and open code that back the claims: from a research paper to an app people use every day.

Loqira Labs

Loqira Labs

An independent edge-AI lab I founded. The idea: AI you own, not AI you rent. Capable models for everyday tasks that run on your hardware, with no cloud and no third-party APIs.

The lab owns the whole stack: original architectures, knowledge transfer, quantization and finetuning, CPU inference in Rust, all taken to shipped products.

loqira.techGitHubHugging Face
01Projects

What has shipped

Product · shipped

KOT

The AI agent harness — one file is all you need

A working AI agent in a single executable for Linux, macOS, and Windows: it reads code, searches the project, edits files, runs commands, remembers what matters between sessions, and drives a team of agents across several providers at once. The web interface is embedded; no sandbox, no skills, no MCP — by design.

  • One executable; web UI, tokenizer, and rules embedded
  • 16 providers: OAuth subscriptions, API keys, local models
  • Sub-agents, delegates, teammates, TOML pipelines
  • Durable memory and full session history on disk
agentkot.ai
Product · shipped

Levitan

On-device voice input: you speak, text appears at your cursor

An app for Windows and macOS: hold a key, speak, release, and the text appears in any input field. Speech is recognized locally; no internet, no GPU, free.

  • Russian, Kazakh, English, and 96 more languages
  • Whisper large-v3-turbo 0.8B: quantized and finetuned
  • A custom Rust inference engine that runs on a CPU
  • ~350 MB, one installer; PolyForm Noncommercial license
LevitanLoqira Labs
Open model

Qwen3.5-0.8B GEC

A text corrector: Kazakh, Russian, English

Open weights under Apache-2.0: the model restores punctuation and paragraphs and fixes typos and spelling. A direct extension of Levitan: voice, then text, then clean text.

  • GGUF ~537 MB; runs in llama.cpp, LM Studio, Ollama
  • ~44 tokens/s on an ordinary CPU
  • Custom 4/8/16-bit QAT with train==deploy parity
  • Apache-2.0: usable in commercial products
Hugging Face
Open source

Data Structure Protocol (DSP)

Graph-based long-term memory for AI coding agents

A skill that gives LLM coding agents durable structural memory of a large repository: a graph of entities, imports, and the reasons behind them, instead of re-reading the code every session.

  • A graph of entities and imports over the whole repository
  • Python; plugs into coding agents as a skill
  • Write-up on Habr
GitHubHabr
Open source

Agent Browser Workspace

A local browser toolkit for AI agents

A toolkit that lets any AI agent drive a real local browser over Chrome CDP and Playwright: research, page reading, and web automation with data staying on your machine. The Tech covered it as a local open-source alternative to Perplexity.

  • Chrome CDP + Playwright, JavaScript
  • Works with any agent; no cloud service in the loop
  • Covered by The Tech and Habr
GitHubHabrThe Tech
Technology stack

The Loqira vertical

From the model's architecture down to the CPU it runs on

Not a wrapper over someone else's API but an owned five-layer vertical, each layer backed by a published work or a shipped product.

  • Own models: MPT-VC, the finetuned Whisper behind Levitan
  • Knowledge transfer: LLM Modules, a capable small model for under $10
  • Context efficiency: Context Merging, attention memory N² → G²
  • Quantization and finetuning for weak hardware, plus CPU inference in Rust
Loqira Labs Research

Discuss a product or an integration

If you want this level of engineering in your product, from local models to automation, write to me.

Write to melibr@bk.ru