Kristopher L. Porath

AI Platform Engineer

Remote (US) · Michigan, Eastern Time · US citizen, no sponsorship required

kitporath@gmail.com · kitporath.com · github.com/kitporath · linkedin.com/in/kitporath

Summary

AI Platform Engineer who builds and operates production private AI infrastructure — and extracts dependable engineering from unreliable models. Architected a production private AI cloud (Kubernetes, GitOps, HA Vault, SSO, internal PKI, self-hosted GPU inference) as a safe substrate for a self-built multi-agent platform, and recovered it from a total power loss with zero data loss. The estate is operated day to day by local agents under my direction. Two decades of operations leadership behind the rigour: up to 35 direct reports, program management, and IT systems administration. Published the failure taxonomy for agent-directed infrastructure work, including my own errors.

Target roles: AI Platform Engineer · AI Infrastructure / LLMOps · Platform Engineer (AI/ML) · Forward-Deployed Engineer

Selected results

• Recovered a full private cloud from total power loss with zero data loss and no split-brain — storage, database quorum, hypervisors and Kubernetes cold-started in dependency order. Executed by an agent under my direction; the only steps requiring me were a BIOS repair at the console and a shamir unseal. Converted into a verified runbook and permanent fixes the same evening.
• Chaos-tested control-plane loss by hard-killing a host running two of three masters. Cluster failed safe, quorum held, and it exposed an anti-affinity gap I then closed.
• Ran a supply-chain campaign with a local model on owned GPUs — 210 → 151 critical CVEs, measured rather than reported. Caught three of the agent's own reporting errors, including 52 images retired as "unused" of which 15 were running the cluster.
• Upgraded RKE2 and Cilium (1.18 → 1.19.4) across six nodes, eliminating 24 critical CVEs with zero workload outage.
• Published an error analysis of agent-directed infrastructure work — 13 audited claims, 8 material errors, and the countermeasures now enforced in CI.

Core skills

• Platform & orchestration: Kubernetes (RKE2), ArgoCD app-of-apps GitOps, OpenNebula/KVM, OneKE, Proxmox, Gitea Actions CI/CD, per-PR environments
• Infrastructure as code: Terraform, Ansible, declarative catalogs exported from live state, drift detection wired to alerts
• Security & identity: Keycloak OIDC/SSO, HashiCorp Vault (3-node HA, transit auto-unseal), cert-manager and internal PKI, Cilium network policy, Falco runtime detection, airgap and egress control, digest-verified image mirroring
• AI / LLM systems: self-hosted serving (llama.cpp), OpenAI-compatible gateways with routing and cost attribution (LiteLLM), GPU and model fleet lifecycle, MCP tooling, multi-agent orchestration with mechanical verification
• Observability: Prometheus, Grafana, Loki/Promtail, blackbox and textfile exporters, alert rules verified firable before shipping
• Storage & networking: LINSTOR/DRBD, MinIO (S3), ZFS, 3-2-1 backups, DNS-as-code
• Languages: Go (AI-directed), Python, Bash, YAML/HCL

AI harnesses & models

Effectively model-agnostic and harness-agnostic — the skill that transfers is the direction and the verification, not familiarity with one vendor's tooling.

• Harnesses: Claude Code, Codex, opencode (cloud and local), Grok, Antigravity, Hermes, and Hirdforge — which I built.
• Frontier models: Claude (Opus 5, 4.8–4.6, Sonnet 5, Fable 5), GPT 5.1–5.5, Gemini, Grok 4.5, Kimi 2.5/2.7, MiniMax 2.5–3, Nemotron Ultra 3.
• Served on my own GPUs: Qwen 3.5/3.6 (27B, 35B-A3B, 122B), Gemma (12B, 31B, 26B-A4B), Nemotron 3 Plus, Step 2.7 and DeepSeek V4 Flash on mixed CPU/GPU, and Laguna S 2.1 at 75GB RAM-resident.

Experience

Sovereignty Labs — Founder & AI Platform Architect

2025 – Present · Remote

• Architected and operate KWS, a production private AI cloud: 19 monitored hosts, 6-node RKE2 Kubernetes (~104 pods), 8 auto-synced ArgoCD applications, 3-node HA Vault with transit auto-unseal, Keycloak SSO, internal PKI, and a digest-verified image mirror for airgapped operation behind deny-all egress.
• Directed AI implementation of Hirdforge, a multi-agent engineering platform: ~67,000 lines of Go across 225 files and 712 merge commits, deployed to production Kubernetes. Built deterministic routing and mechanical CI gates so agent output is verified by tests rather than model self-assessment.
• Built and operate a self-hosted GPU inference layer serving OpenAI-compatible APIs across a multi-GPU fleet, with a gateway providing per-role routing, quotas and cost attribution.
• Made the estate operable by agents: the runbook, 50 dated incident records and 39 proven primitives are the corpus agents read to learn how the system works, with live state always retrieved by tool call rather than from documentation.

Prefix Corporation — Manufacturing Operations & IT

2010 – Present · Michigan

• Automation Logistics (2024–Present): material logistics for an automated industrial paint line; management-track role.
• IT System Administrator (2023): administered company IT systems — the point where self-taught infrastructure work became the day job.
• Program Manager (2020–2021): cross-functional program and delivery ownership.
• Team Lead → Supervisor (2013–2022): led up to 35 employees; ran production operations for quality, throughput and on-time delivery against fixed schedules.

Saleen Inc. — Team Lead

to 2008 · Michigan

High-performance automotive finishing.

Certifications & learning

• Google Cybersecurity Certificate; TryHackMe.
• Self-directed and hands-on, evidenced by four generations of production systems — Valhalla → Asgard → KWS → Hirdforge — rather than coursework.