← Selected work
Platform infrastructure 2026

Playground

A private server platform built one layer at a time, with each layer tested, documented, and proven to survive a reboot before the next one starts.

Active / commissioned in layers

The idea

The discipline around the stack is the point: establish a baseline, change one layer, prove it, document it, and prove that it can recover.
Playground — public system view
01 Compute Virtualized, bounded substrate
02 Platform Ingress, data, and service layers
03 Observe Health, metrics, and negative tests
04 Recover Restore and reboot proof
05 Record Runbooks and service handoffs

A lab with acceptance criteria

Playground is a private infrastructure environment built to make platform engineering concrete. Services are introduced in bounded tranches with explicit prerequisites, resource limits, network boundaries, acceptance checks, and a durable operating record.

Reliability is part of the feature

Backup coverage is inspected, recovery copies are validated, restore paths are exercised, and critical layers are checked again after reboot. Each failure becomes a revision to the method and the operating record.

Designed for another operator

Each commissioned service leaves behind a record of purpose, dependencies, boundaries, recovery behavior, and the evidence used to accept it. The system is not considered complete if only its builder can understand it.

What this work demonstrates

Evidence, not a capability list.

  • Tranche-gated commissioning.

    Services go live in bounded groups. Each group has written prerequisites, resource caps, network boundaries, acceptance checks, and recovery evidence before the next group starts.

  • A backup that lied.

    A nightly pull path could report success without proving the transfer. The coverage audit that exposed the gap now distinguishes a clean run, acknowledged exceptions, and unhandled failures.

  • Restore, not backup.

    Recent checks have exercised PostgreSQL and Qdrant restores, TLS state, and service recovery after reboot. Recovery copies are opened and tested instead of merely counted.

  • Reboot is an acceptance test.

    A service is accepted only after it returns healthy on its own following a host restart. The check is part of commissioning, not a future hardening task.

  • Pull-only by design.

    The backup host reaches in and takes what it needs. The protected host holds no credentials for the backup destination, limiting what a compromise on either side can reach.

Public boundary

Public diagrams deliberately describe layers and operating method, not the private environment's addresses, topology, credentials, access rules, or recovery destinations.