RKFleming

Forward Deployed Platform Engineer

Ryan Fleming

I ship platforms where the operators already are.

Cutovers, clusters, and identity at the edge of production — embed with the team, scale for the peak, leave systems that still work when I roll off.

Now Agentic platforms by day · a 180B model on a desktop APU by night

Selected work Open a session
1,500+
nodes at peak · RKE2 / Karpenter
1,000+
sites · Harvester HCI footprint
100K+
identities managed · career
28K
users · Entra federation cutover
262k
token context · local 180B model

02 — Skills

Tools I keep sharp.

Marks from the stack I run in production — the ones with logos. Everything else I work in sits below.

Kubernetes
RKE2
Rancher
Longhorn
Helm
Docker
Terraform
Grafana
Prometheus
ClickHouse
Metabase
Cloudflare
nginx
GitLab
GitHub
Linux
Red Hat
OpenShift
Okta
FortiGate
Cisco
Go
Python
Rust
Svelte

Beyond the logos

Compute & orchestration

  • Karpenter
  • Harvester HCI
  • KubeVirt

Identity

  • Microsoft Entra ID
  • Active Directory
  • OneLogin
  • Conditional Access

Network & endpoint

  • Meraki
  • CrowdStrike
  • Intune
  • NinjaOne

Observability

  • Alertmanager

Languages

  • PowerShell
  • KQL / SQL

Cloud & ops

  • Azure / Arc
  • AWS
  • cloud-init
  • ServiceNow

AI & inference

  • llama.cpp
  • Vulkan / ROCm
  • GGUF
  • OpenCode
  • Qwen3.8
  • MTP spec-decode

03 — Work

Field work that shipped.

Outcome-first case studies — the problem on the ground, the path taken, and the number that stuck. Deployments and the AI work, in one place.

01 Scale
1,500+ nodes at peak

RKE2 + Karpenter that breathes with the day

Karpenter grows the RKE2 pool into the rush, then releases capacity when demand falls.

Challenge

Demand spikes hard during peak operating hours. A fixed node pool either burns money idle or runs out of headroom when the field gets busy.

Approach

Stood up and ran RKE2 clusters with Karpenter so capacity follows real demand — nodes provision and drain with the operating day, not a static reservation. Tuned disruption budgets and node pools with the operators who live the rush.

Impact

Clusters expand toward 1,500+ nodes when the peak hits, then contract when demand falls — without babysitting the fleet by hand.

  • RKE2
  • Karpenter
  • Autoscaling
  • Peak capacity
02 Platform
1,000+ sites migrated

VMware to Harvester without taking sites offline

Challenge

Hypervisor sprawl across a national footprint. Sites could not go dark for a migration window.

Approach

Embedded with site ops to consolidate onto Harvester HCI — VM tuning, Longhorn storage, and a cutover path that kept workloads online through the move.

Impact

Sprawl collapsed onto one HCI stack. Cutover stayed boring for the field.

  • Harvester
  • HCI
  • Longhorn
  • VM migration
03 Identity
28K users cut over

OneLogin and Okta → Entra, phased for the field

Challenge

Legacy federation still held roughly 28,000 users. A hard MFA cliff would have lit up every helpdesk queue.

Approach

Sat with identity and helpdesk on a phased sunset — conditional access by population, MFA that landed quietly, and a cross-tenant federation prototype along the way.

Impact

Federation cut over with almost no field noise.

  • Entra ID
  • Okta
  • Conditional Access
04 Side project
Go DR product

Resilium · failkit.io

Challenge

Harvester-to-AWS and mixed-hypervisor shops had no clean DR orchestration product.

Approach

Built the product I wished existed at work — Go services that orchestrate disaster recovery across heterogeneous hypervisors.

Impact

A working DR path for shops that do not fit a single-cloud playbook.

  • Go
  • DR orchestration
  • SaaS
05 Foundations
20K SaaS customers

National SaaS, K-12, and the shop floor

Challenge

From RDS farms for a national software vendor to district identity and Fanuc/Mitsubishi machine telemetry — places where uptime is personal.

Approach

Delivered multi-tenant session fleets, ran identity and endpoints for education, and led an OEE pipeline into live floor visibility.

Impact

Scale habits forged where failure shows up as people waiting — still how I ship.

  • RDS
  • Identity
  • OEE
  • OT
06 AI · Home lab
180B MoE · 262k ctx · self-hosted

A frontier-class model, running in the house

2 slots45.732.510.57.025.821.515.49.4×1×2×4×8spec-decode onspec-decode offtok/s per requestrequests at once
Speculative decoding makes a single session fly; past two concurrent requests it works against you — so the box runs two slots: one fast session plus room for a subagent.

Challenge

Renting cloud GPUs puts a meter on every token and hands the prompts to someone else. Wanted a large open model available to agents offline — private, always on, fully owned.

Approach

A compact desktop AI machine — GMKtec EVO-X3, AMD Ryzen AI Max+ 395 with the Radeon 8060S iGPU, 128 GB unified memory — serves Qwen3.8-Flash-Next, a ~180B-parameter mixture-of-experts model with a 262k-token context, through llama.cpp built on Vulkan. Exposed as an OpenAI-compatible API as a boot-time service. Quantized to IQ4_XS so the weights plus a long context fit in unified memory; MTP speculative decoding for speed.

Impact

A 180B model answering locally at roughly 37 tokens/sec, up to ~45 on code. Found a fact buried in a 161k-token prompt, then re-answered from cache in about a second. Same model the coding agents call.

  • llama.cpp
  • Vulkan
  • GGUF
  • Qwen3.8
  • MTP spec-decode
07 AI · Practice
Multi-agent how the work ships

Building with agents, not just about them

Challenge

Agentic coding tools move fast and drift faster without guardrails, scope, and a human who knows the production context.

Approach

Runs a multi-agent workflow — a lead agent plans and orchestrates, scoped subagents do parallel, file-partitioned work, everything lands through GitOps and CI. This very site builds multi-arch images on tag and rolls out through a Fleet GitOps bundle. Humans set direction and review; agents run the loop.

Impact

Work you can watch get done — this portfolio section itself was drafted by a pair of subagents and shipped through the same pipeline it describes.

  • OpenCode
  • Multi-agent
  • GitOps
  • SvelteKit

03 — Approach

How I work in the field.

  1. 01

    Sit with the operators

    Discovery happens on the floor and in the NOC, not only in a slide deck. I learn the real failure modes before I touch the cutover plan.

  2. 02

    Ship without taking the site down

    Migrations, MFA, and platform swaps land in phases. Production keeps running while the new path takes traffic.

  3. 03

    Scale for the peak, not the average

    Capacity should follow the operating day. Clusters that grow into the rush and shrink when it ends — Karpenter and friends, not a permanent oversized pool.

  4. 04

    Leave metrics and runbooks behind

    When I roll off, the pager still has a signal and the team has a path. Observability and docs are part of the deliverable.

05 — Off the clock

When the pager is quiet.

Diesel & overland

Built Ram 2500 Cummins — G56 manual swap, Carli suspension, 37s. Overland when the pager is quiet.

Land Cruisers

Dixieland Cruisers — Toyota Land Cruiser club, organized and run as a 501(c)(3).

HF radio & SDR

Chasing FT8 across the bands and learning the waterfall. SDR hardware on the bench.

Rivers & ovals

River floats with GPX in hand, and NASCAR on the weekend. Data follows me everywhere.

06 — Contact

Open a session.

Forward deployed platform engineering · Tuscaloosa, AL

[email protected]
LinkedInGitHub