Stockholm, Sweden / Platform engineering

ROMAN
USOV.

Senior Platform Engineer at Pierce

From the physical rack to AI operations.
I follow the problem through every layer, then leave behind a system other people can run.

Three investigations. One way of thinking.
The whole column.Compute → Network → Cloud → AI
AT PIERCE / 24MX · XLMOTO · Sledstore15+ years in technology10+ years owning production infrastructure
01 / Pierce · Network source of truth

The network nobody
could describe.

Documents existed. A reliable answer to “what breaks if I touch this port?” did not.

Knowledge of the estate
A network made traceableConceptual comparison: disconnected knowledge in diagrams, memory and configurations becomes a NetBox model of VLANs, cables and ports, supported by switch configuration backups. Not a production topology.Old diagramsSwitch configsRecollectionRack visitsWhat is true?NetBoxONE SOURCE OF TRUTHVLANsCablesPortsTraceable relationshipsSwitch configuration backups
Conceptual model, not the production topology.

I traced the estate and modelled it in NetBox. VLAN by VLAN. Cable by cable. Port by port.

The picture had lived in people's heads, old diagrams and drifting switch configurations. I made it something the next engineer could query, with configuration backups behind the switch fleet.

Now we can reason about it
before we touch it.

Faults can be traced. Changes can be reviewed. Improvements can be proposed.

NetBoxNetwork modellingConfiguration backups
02 / Pierce · An independent investigation

Same URL.
Same 200.
Different reality.

Independent audits looked bad. Internal reporting looked healthy. I compared the raw origin responses by user agent.

Recognised crawler200 OK
Navigation. Products. Copy. Links. A fully rendered page.
Other user agents / cold cache200 OK
EMPTY.A shell without the page content
The same origin returned an empty shell to everything outside its recognised list.

Conceptual response comparison. No production page or response body is reproduced.

The method

Read what the server actually returned. Compare the bodies, not just the status codes or keyword dashboards.

The finding

The underperformance had an infrastructure cause. A commercially important defect had survived between teams' remits.

03 / Pierce · Linux fleet

Different estate.
Same missing truth.

Then I applied the same pattern to packages and patching: Red Hat Satellite and Pulp as the source of truth for what exists on which machine.

Content managed deliberately. Patching that can be audited. A stepwise platform upgrade, with the fleet readied for its next operating systems.

Red Hat Satellite / Stepwise upgrade
6.136.19
ModelPackages and repositories
ManageContent across the fleet
AuditPatch state per machine
Readied for RHEL 10Readied for Rocky 10
Pierce / Further outcomes

Work that stays working.

The test is what remains after the change: known state, repeatable operations and fewer things that depend on memory.

3 TB

Monitoring dataset rescued by database tuning.

Pierce
0

Certificate-expiry outages since automating the lifecycle.

Pierce · Since automation

Sole infrastructure on-call through Black Week, seven days straight.

Pierce · 2024 & 2025
1 estate

Migrated from Akamai to Cloudflare, including edge routing and Terraform reconciliation.

Pierce · 2026
Scope / Physical to operational

The whole column.

Depth at one layer helps. Following a problem across the boundaries is where I do my best work.

08

AI operations

AT PIERCE

Internal LLM service deployed company-wide.

INDEPENDENT INFRASTRUCTURE

Multi-agent operations, persistent memory and scheduled health rounds.

07

Cloud

AT PIERCE

AWS organization built from the root account up. Entra ID SCIM, dual-active IPSec design, Azure and Cloudflare Workers.

06

Platform

AT PIERCE

Docker and Harbor. Harvester 1.7 / RKE2 proof of concept on nested vSphere. API-gateway evaluation and ADR: Kong against KrakenD.

INDEPENDENT INFRASTRUCTURE

Talos Kubernetes, Cilium and Kube-OVN.

05

Observability

AT PIERCE

Zabbix 7.0 LTS with automated proxy deployment. ELK major upgrade, Grafana, Kafka consumer lag via Burrow and node_exporter.

INDEPENDENT INFRASTRUCTURE

VictoriaMetrics.

04

Edge & mail

AT PIERCE

Akamai to Cloudflare migration, Workers routing and Terraform state reconciled with production. SPF, DKIM and DMARC across all domains, with SPF flattening.

1,363 pages untangled in a category migration; over 1,300 retired campaign redirects cleared across two markets.

03

Network

AT PIERCE

NetBox modelling, switch configuration backups, firewall changes by API, VLANs and IPSec/VPN.

INDEPENDENT INFRASTRUCTURE

VyOS routing.

02

Storage

AT PIERCE

NetApp NFS/CIFS. Blade-controller latency traced to NFS nconnect tuning.

INDEPENDENT INFRASTRUCTURE

LINSTOR / DRBD and ZFS.

01

Compute

AT PIERCE

VMware / ESXi on Cisco UCS, Linux, Red Hat Satellite and Pulp.

INDEPENDENT INFRASTRUCTURE

Proxmox.

Build it so the next person
does not need you in the room.

The secrets and certificate platform follows that rule: deployments pull secrets at deploy time; certificates renew and deploy themselves. Rotation becomes a repeatable process, rather than a favour asked of the person who built it.

Independent / Built in my own time

After the day job.

Different constraints. The same engineering standards.

01 / Own infrastructure

An operations layer
for a team of one.

Talos Kubernetes on Proxmox. Cilium and Kube-OVN networking, LINSTOR/DRBD and ZFS storage, VyOS routing.

The constraint is one person and no rota. The operations workflow is designed to turn collected state into a decision about what needs human attention.

Scheduled agent workflow
  1. Collect state across the layers
  2. Compare with the previous report
  3. Triage what changed
  4. Decide whether a person is needed
  5. Page with the evidence attached
From an item to a listing
OWN PRODUCT / LIVE

Uselist

A Telegram-first inventory and listing assistant. Photograph an item, get recognition and a ready-to-post listing, publish to marketplaces.

In production

Real users, Stripe billing and eBay integration. Python, Telegram Bot API, PostHog and Grafana.

A living body of knowledge
OWN PROJECT / NON-COMMERCIAL

Batrachos

A Ukrainian biology and ecology portal. Migrated from Drupal to Django and Wagtail through a rehearsed cutover, with no acceptable downtime window.

Visit Batrachos
Gear, movement and upkeep
OWN PRODUCT / WORKING ALPHA

Layrova

A kit and upkeep companion for cyclists and runners. Photo-based gear cataloguing, weather-aware clothing and component wear from ride data.

Current stage

A working alpha. An ongoing product, not a finished release.

Roman Usov / Stockholm, Sweden

Good systems start
with good questions.

English · Ukrainian
Swedish (improving)