Amir Salmani

Platform & infrastructure engineer · Tampere, Finland

Everything you ship runs on someone’s infrastructure. I build that part.

Self-managed Kubernetes on bare metal and regional clouds — clusters, storage, networking, and the delivery and observability that keep them honest. I’d rather own the whole stack than rent it, and I build it to be handed over.

Status
Open to roles
Based
Tampere, Finland · UTC+3
Looking for
Senior platform / SRE · remote across Europe
Also
Contract & fractional · EU invoicing
01

The stack, top to bottom

select a layer

L5 → L1 · abstraction down to metal

02

Receipts

production, not demos
  • 31 servers across two datacentres Platform + DevOps function built from zero
  • < 50 working days from nothing to live 80-person AI company · no prior platform team
  • 1 datacentre lost, cluster rebuilt Talos + Cilium · disaster recovery, for real
  • RPO 0 data loss, 2–3 min automated failover Sync-replicated Postgres · 472-user platform
  • 700,603 Flink checkpoints, zero failures 8 months live · Kafka consumer lag held at 0
  • −7.3 h of downtime removed per month 99% → 99.99% SLA · Docker Swarm → RKE2
  • 20 min → 3 deploy cycle, and 5–8 min merge-to-live GitLab CI/CD, then Flux with digest-pinned images
  • 4× H100 GPU inference in production SGLang · 30k+ requests/day · tested to ~230k images/day
03

AI in the loop

built, not subscribed

Most engineers use an AI assistant. I build the parts it plugs into — so an agent works against my actual infrastructure records, and cannot walk out of the door with a client’s material while doing it.

  • MCP server · self-built obsidian-zed-mcp-server

    Exposes a structured engineering vault over the Model Context Protocol, so an agent queries real runbooks, incident records and architecture decisions instead of guessing at them.

    Any MCP client · read and write
  • Agent skill · self-built data-guardian

    A scrub gate. Client names, internal URLs, host names and credentials do not leave the vault — nothing reaches a public surface until it has cleared.

    Why sanitised case studies can exist at all
  • Agent skills · self-built second-brain
    web-clip-and-file

    Capture and filing, so a design decision or a 2 a.m. fix gets written down while it is still true, rather than reconstructed from memory months later.

    Johnny.Decimal vault · 663 notes and counting

Why this is on the page at all: it is the reason every figure above carries a source line instead of a claim. The records exist because capturing them is part of how I work, not something done afterwards for a CV.