Local inference
Production LLM serving on your hardware: llama.cpp for CPU and mixed fleets, vLLM for high-throughput GPU nodes. Quantized, monitored, versioned.
- llama.cpp
- vLLM
K109 — full-cycle engineering bureau
online · all systems nominal
From architecture and UI/UX to hardened operations — and AI that never leaves your perimeter.
Six service lines — from product development to AI engineering.
We build products across the whole surface: React or Angular SPAs, Astro and Next.js sites, NestJS services, React Native apps, Electron desktops. One team, one responsibility — from first mockup to production.
What we deliver
Infrastructure described in Terraform/OpenTofu, configured with Ansible, delivered through GitOps pipelines on Kubernetes and Argo CD. Observability is built in from day one: Prometheus, Grafana, Loki, Jaeger.
What we deliver
The full test pyramid in CI: unit and integration with Vitest and Testing Library, E2E with Playwright and Cypress, load with k6, accessibility with axe-core. Every merge passes the gates your users depend on.
What we deliver
SAST/DAST and OWASP ZAP scans in pipelines, SSO on Keycloak and LDAP directories, hardened network design and backup schemes with Bareos — verified by real restores, not by hope.
What we deliver
C4 maps for every system level, an honest choice between monolith and microservices, domain-driven design backed by patterns — and UI/UX in open-source Penpot, readable even by AI agents.
What we deliver
Local inference on llama.cpp and vLLM, RAG over your documents on Qdrant, agent-assisted delivery with Cline and OpenCode, automation with n8n and Flowise. No third-party APIs, no data leaving.
What we deliver
70+ tools across QA, DevOps, development, security and operations.
Architecture first: C4, domain patterns, shift-left quality, CI/CD automation.
Every engagement starts with a map, not a framework. We model the system at four C4 levels so business, architects and developers argue about the same picture.
Architecture by fitness, not fashion. Most systems should start as modular monoliths; distribution is a cost we pay only when it earns its keep.
Patterns compress decisions. Classic design keeps code changeable; distributed patterns keep systems honest under failure.
Design
Distributed systems
Defects get cheaper the earlier they die. Quality gates stand at every step of the path, and every gate runs automatically.
Every change walks the same path. Code that was not deployed by the pipeline was never really tested at all.
Local LLM serving, RAG on Qdrant, agent automation — data never leaves your perimeter.
Production LLM serving on your hardware: llama.cpp for CPU and mixed fleets, vLLM for high-throughput GPU nodes. Quantized, monitored, versioned.
Your documents, embedded and stored in Qdrant inside the perimeter. Answers come with sources — corporate knowledge without corporate leaks.
Coding agents Cline, KiloCode and OpenCode for delivery teams; n8n and Flowise pipelines for the rest of the company. Humans approve, agents grind.
SAST/DAST, penetration testing, SSO, backups, hardened networks.
Transparent delivery: Redmine, Zulip, BBB, RustDesk, Asterisk, mailcow.
Delivery tooling runs on our own servers — your project data never touches third-party SaaS.
Every task, deadline and decision is visible to the client.
Decisions live in threads; meetings are recorded and minuted.
Hands-on help: screen sharing, voice, on-site visits.
Reports and notices come from our own mail, on our domain.
An engineering bureau: full cycle, deep ownership, no hand-offs.
K109 reads like a design bureau designation. That is the idea.
We are a full-cycle engineering bureau: architecture, code, platform, quality, security, operations — one team owns the outcome end to end. Drawings before code, tests before release, restores before trust. Boring technology, sharp engineering, and AI that stays inside your perimeter.
Tell us about your system — we answer within one business day.
An engineer will reply within one business day.