DK
TR

About

The long version

The short version lives on the home page. This is the long one — how I work, what I have built, and the road here.

Focus

Hardware first. Everything else follows.

Two decades of enterprise infrastructure, pointed at one question: how do you make AI run reliably on hardware you actually control — small, cold, power-starved, or a hundred million kilometres away?

A satellite in orbit — compute at the edge of the atmosphere

Now · Cosmic DynamiX

Onboard compute for space

AI Hardware Engineer — with the core team since the company was founded in 2026.

Cosmic DynamiX builds low-power computing hardware and software that let spacecraft process data before transmitting it — more science per bit of downlink.

  • Hardware strategy — evaluating low-power chip architectures for maximum energy efficiency.
  • Performance optimization — tuning AI models for specialized hardware and mission-critical reliability.
  • Edge intelligence — hardware-level optimization for autonomous processing beyond Earth's orbit.

Since 2023 · Local AI & LLM

Running AI on hardware you own

Through kumova.net — AI-driven infrastructure and hardware optimization.

Local LLM environments engineered as one problem: silicon, GPU allocation and model tuning, measured end to end in tokens per second and latency.

  • Local LLM deployment — designed for token-per-second throughput and low-latency inference.
  • GPU orchestration — vGPU and passthrough for private model training and deployment.
  • Quantized models at the edge — energy-efficient deployment strategies for edge devices.
  • Systems integration — Docker, Proxmox and AI accelerators wired into existing enterprise infrastructure.

Expertise

Twenty-three years in the machine room

Before AI became the headline, the job was keeping critical systems alive — networks, broadcasts, storage, telephony, platforms. That discipline is what AI hardware work stands on.

Rows of servers in a data centre
Infrastructure architecture
Scalable network and systems design, complex hardware rollouts and 99.9% availability for mission-critical operations.
Live broadcasting & VoIP
High-capacity internet-based live broadcast systems and custom VoIP hubs — engineered, integrated and operated end to end.
CDN & high availability
Server-side architecture, performance scaling and content delivery for large-scale digital platforms (incl. Billur.TV).
Virtualization & orchestration
VMware ESX/vSphere, Proxmox, Docker — vGPU and passthrough topologies for private AI workloads.
AI hardware & LLM ops
Fine-tuning, quantization, RAG pipelines — and getting real tokens per second out of constrained hardware.
Storage & self-hosting
Synology, QNAP, FreeNAS/TrueNAS — self-hosted services, backup regimes and hardware lifecycle management.
Digital platforms & publishing
Drupal, Joomla, WordPress, OpenCart, Xibo — corporate platforms, e-publishing and digital signage at scale.
Networks & core systems
Active Directory, Windows Server, CentOS/Linux, mail and telephony — the unglamorous layer everything depends on.

Selected work

Measured, not assumed

Representative engagements from the last three years of AI hardware and infrastructure work.

Making measurements on the bench

01

Local LLM inference rig

A private environment for large language models where throughput is measured, not assumed. Hardware selection, GPU allocation and model tuning treated as a single optimisation loop.

— Tuned for token-per-second throughput and low-latency inference.

LLM INFRA · TOKEN/S · LOW LATENCY

02

Private AI on GPU

vGPU and passthrough orchestration so models can be trained and served without data leaving the building — data privacy as an architecture decision, not a policy paper.

— Cost-efficient private training and deployment on owned hardware.

VGPU · PASSTHROUGH · DATA PRIVACY

03

Quantized models at the edge

Shrinking models until they fit the power budget of edge hardware — then proving they still behave reliably under mission constraints.

— Deployment strategy for energy-efficient autonomous processing.

QUANTIZATION · EDGE · POWER BUDGET

04

Low-power AI accelerator evaluation

Benchmarking low-power AI accelerator prototypes kernel by kernel: decode throughput, attention operators, watts per token. Silicon Valley lab testing, 2026.

— Numbers that decide whether silicon is worth flying.

BENCHMARKS · SILICON · WATT/TOKEN

Client and vendor names are held back here; full references available on request.

Timeline

2000 → today

From print layout to spacecraft compute.

Generations of computing machinery

  1. 2026 →

    AI Hardware Engineer · Cosmic DynamiX · Cupertino, CA

    On the core team since the company was founded — low-power compute for spacecraft, from chip evaluation to edge intelligence.

  2. 2023

    Local LLM & AI hardware · AI turn

    Working deeply on AI begins: local LLM infrastructure, token-per-second optimization, quantized models on energy-efficient hardware.

  3. 2019 – 2026

    Founder & consultant · kumova.net

    Consultancy on AI-driven infrastructure and hardware optimization — local LLM environments, GPU orchestration, low-power edge deployment. Technical backbone for community platforms such as Turkish-American TV and Kozmik Anafor.

  4. 2010s

    Infrastructure & CDN Solutions Architect · Billur.TV

    High-availability infrastructure and content delivery for large-scale digital platforms.

  5. 2003 – 2019

    Information Technology Specialist · TÜSİAD, İstanbul

    ~16 years owning mission-critical IT for one of Türkiye's leading business associations: infrastructure architecture, internet-based live broadcasting, VoIP, hardware lifecycle, digital transformation projects and the technical backend of large corporate platforms.

  6. 2003

    Production director · Ulusal Kanal (TV)

    Live production workflows and broadcast operations.

  7. 2000

    Graphic designer · Prestige Publishing

    Print layout and visual craft — where the eye came from.

Off the clock

The other half of the craft

AI hardware by day, frames, film and notes by night — all of it from the same curiosity.

Hands on the piano keys

2025

Cinematography
Gülden — "Yan Ulan Yan" music video, shot on the streets of New York (dir. İzzet Başlak).

2025

Photography
"Moments of Istanbul" concert, Manhattan School of Music — with Adem Merter Birson.

2010

Stage
Adım Tiyatro — "Yolumuzu Bulalım", European tour.

2009

Sound design
Feature film "Bozkırda Bir Çiçek" (dir. Hasan Karcı).

ONGOING

Keys & music
Keyboard player. Instrumental arrangements on YouTube.

COMMUNITY

Giving back the stack
Web and infrastructure support for Turkish-American TV; technical backbone of the Kozmik Anafor astronomy platform.