Andaç Kartal

Independent IT consulting · Istanbul, Türkiye

I take on the infrastructure work there is never time for.

scoped by a human · executed by agents · logged end to end

Patch backlogs, migrations, hardening, the runbook nobody has time to write. I scope the work and stay accountable for it. Agents carry out the execution over the same interfaces your administrators already use — WinRM, SSH, and your vendors' APIs — and every step lands in a log you can read afterwards.

Send me a task →

01

Four hypervisors, three countries

This is not one client's estate. It is the range I work across day to day: enterprise vSphere on an HCI cluster, my own Proxmox node, servers in two clouds on two CPU architectures, and the container layers running on top of all of it. Every figure below was measured on 26 August 2026 — queried from the platform itself, not remembered.

platforms

hypervisors
VMware vSphere 7 · Proxmox VE 8 · Hetzner Cloud · Oracle Cloud (ARM64)
container layers
Docker · Incus/LXD system containers · Coolify PaaS
virtual desktops
VMware Horizon · ~50 desktops · ~31 concurrent sessions
network and identity
pfSense — itself a VM · WireGuard mesh · HAProxy · Active Directory
storage and backup
HPE SimpliVity HCI · ZFS · LVM-thin · Veeam · Proxmox Backup Server · S3

scale measured, not rounded

physical hosts
5 · 128 cores / 256 threads · ~1.37 TB RAM
virtual machines
99 across four hypervisors · 71 on vSphere alone
containers
19 system (LXC/Incus) · 88 application (Docker)
production storage
~18 TB verified SAN · 3.4 TB on-prem pools · ~6 TB file services
backup estate
49 TB NAS repository · 10 TB offsite object storage · 191 snapshots
protected daily
25 VMs with a restore point in the last seven days
private network
29 nodes · 3 countries · x86_64 and ARM64 · no open management ports
verified access
a proven method per host — an open port is not the same as a manageable one

beyond the hypervisor

inventory of record
9,477 assets · 157 managed endpoints · 34 thin clients
written memory
94 documents · 15,100 lines · 26 recorded failure patterns
money-holding system
3,162 wallets · append-only ledger · 113,766-row reconciliation
delivery cadence
2,335 commits · 216 migrations · 7 months unbroken
field experience
22 years · systems and infrastructure in manufacturing

02

How an engagement runs

what you and I decidewhat the agents do

  1. 01

    Scope

    We agree what may change and what must not be touched. That boundary is the engagement — not a preference, a limit.

    boundary compiled into pre-flight checks; the run refuses to start until they pass

  2. 02

    Runbook

    I write the procedure and walk you through it before anything touches a host. You see the commands, not a summary of them.

    procedure executes against the hosts you named, over WinRM, SSH, or vendor APIs

  3. 03

    Exceptions

    Anything the runbook did not anticipate comes to me. Destructive steps wait for your approval, every time.

    an out-of-scope condition halts the run and reports it — never improvises around it

  4. 04

    Evidence

    You finish with a record you can hand to an auditor, and a runbook your own team can run again without me.

    every command, result and timestamp written to the run log as it happens

03

Three pieces of work, described in numbers

Client names are withheld. The figures are here to show the scale of the work, not the state of anyone's estate.

  1. A · manufacturing plant · ongoing

    Making a production estate handover-ready

    A production estate grown over two decades: ~24 VMs, 9,477 inventoried assets, two vCenters, a 6 TB file server, 11 print queues. The brief was to make all of it operable by someone who had not lived through its history.

    What that took: a canonical card per host and application, a verified access matrix that separates the port is open from we can actually manage it, and a library of 26 failures this environment has actually produced — each with the symptom that gives it away.

    a new technician — or an agent — is productive on day one; a handover becomes a file, not an event

  2. B · dealer networks · in production

    A system that holds other people's money

    3,162 employee wallets across four corporate dealer networks, millions of lira in balances and unissued vouchers. Boundary: no write access to the production database, ever; schema changes only through reviewed, git-first migrations.

    Balances are never set by hand — a trigger writes them, and a correction is a reversing entry, not an edit. The accountant's reconciliation pack, 113,766 rows, is generated read-only from production.

    every balance can be traced back to the entry that produced it — the reconciliation is evidence, not a summary

  3. C · shop floor · shipped

    A dashboard that stopped polling

    A CNC machine-status board polled the MES database continuously to stay current. The database is production, and read-only was not negotiable.

    Rebuilt as a single reader pushing over WebSocket, with the polling loop removed entirely.

    ~0.3% CPU and roughly 99% less bandwidth, at the same real time

04

What I take on

05

What I won't do

If you are letting agents near production, this list matters more than the one above.

06

How we would work together

engagement
A monthly retainer against a written scope. What is inside is guaranteed; what is outside is quoted separately before it runs.
boundaries
Before anything runs we agree what may change and what must not be touched. That list is the engagement.
exit
Thirty days' notice, either direction. The documentation stays with you. No lock-in — that is the point of writing it all down.
knowledge transfer
Training your people is a separate, time-boxed, billed item. Never a quiet giveaway, and never held hostage.
the paperwork
Invoiced from a Turkish sole proprietorship, VAT added on top. Correspondence in English or Turkish.

07

Start small

a defined first step

A two-week infrastructure review. You end up with a written account of what you are actually running, where it is exposed, and what to fix first — in the order I would fix it.

Read-only throughout. Nothing changes on your systems during the review, and no commitment is implied beyond those two weeks.

Or pick one task from the queue and send it to me. Scope, price and the limits we agree on come back in writing before anything runs.

andac@andackartal.com

Correspondence in English or Turkish.