Michael Olszewski

PharmD, BCPS, BCCCP · DevOps Engineer · Clinical AI Developer · Charlotte, NC molszewski423@gmail.com

Twenty years in critical care and infectious disease gives you a particular perspective on clinical data - you learn quickly that the difference between a real safety signal and statistical noise can matter enormously. That experience is what led me to build my own clinical AI tooling rather than wait for commercial platforms to catch up. I led my organization to IDSA Antimicrobial Stewardship Center of Excellence designation, and somewhere along the way, a decades-long passion for Linux became the infrastructure that runs it all.

Everything here - the clinical AI platforms, the homelab infrastructure, the DevOps and cloud work - has been built entirely in my own time, outside of my traditional clinical pharmacist role. That is precisely the point. This is what I do because I find it genuinely compelling, and it is why I am actively looking for consulting and freelance opportunities where these skills can be put to use.

The technical foundation goes back further than most. At 10 years old I was writing BASIC, cracking open computers to learn how the hardware worked, and dialing into BBS services on a 286 running MS-DOS - years before the web existed. The command line, the hardware, the problem-solving at the system level - none of it has ever felt foreign. What has changed is that four decades of accumulated instinct now has a real and productive outlet.

LocumView

locumview

Open-source VDI platform on Red Hat Enterprise Linux 10, and my most direct platform-engineering portfolio piece: hardened GNOME/Sway desktops provisioned by Terraform and Ansible, delivered through the browser via Apache Guacamole behind Keycloak SSO and MFA, RDP never exposed, every environment rebuildable from the repo. STIG/CIS hardened with OpenSCAP evidence committed alongside the code, mapped to HIPAA technical safeguards.

Beyond the base desktop, it's a platform for research and local LLM usage in regulated environments - the same hardened endpoint carrying agentic tooling and generative applications (clinical and research tooling to start, generalizing to other regulated domains later) built on local models, so nothing regulated leaves the boundary.

Status: early build, Phase 1 (hand-built reference desktop) in progress. Full architecture, phase plan, and ADRs live in the repo.

Clinical AI

pv-workbench

Pharmacovigilance platform for signal detection, MedDRA coding, ICSR narrative drafting, and literature monitoring. Built around a tiered LLM stack: gemma4:26b for regulatory Q&A and signal interpretation, gemma4:e4b for ICSR narrative generation, and qwen2.5:7b for fast intent routing. PRR/ROR disproportionality analysis against FDA FAERS using Evans criteria - with continuity correction for sparse data, Yates' chi-squared correction, and artifact exclusion. RAG over ICH/FDA/EMA guidelines via ChromaDB. All inference runs on a local RTX 5060 Ti via Ollama - no patient data leaves the machine. Deployed as a k3s pod at http://pv.lan.

argus-bot - Discord bot (Argus#1432) providing mobile access to pv-workbench for signal queries, case lookups, and narrative review.

ams-intelligence

Antimicrobial stewardship platform surfacing resistance trends, utilization signals, and clinical decision support. Backed by two decades of hands-on AMS program leadership including IDSA Center of Excellence designation. Confounding by indication is explicitly modeled - last-resort antibiotics treat the sickest patients in the building; the platform is built to catch that bias. Deployed at http://ams.lan on the same k3s cluster.

RingCatch Agency

ringcatch-agency (private repo) · ringcatch.io

Early-stage AI chatbot agency for US local SMBs - HVAC, plumbers, electricians, dental, auto repair, law firms. Launched May 2026, pre-revenue, actively in outreach. Pricing: $450 setup + $89/month. Twenty-four services running in Kubernetes across the cluster - the full pipeline from lead scraping through booking is built and operational, working toward first clients.

Built as an autonomous multi-agent system: a FastAPI orchestrator with 22 tool endpoints drives scraping, outreach, inbox monitoring, and billing agents over an internal event bus, with self-healing recovery (a support agent that detects unhealthy pods and restarts them via the Kubernetes API, capturing logs first so nothing is lost) and automated 6-hour health reporting.

LLM routing: Gemini 2.5 Flash → Ollama gemma4:26b (k3s GPU pod on MikePC) → Groq llama-3.3-70b → Groq llama-3.1-8b

Public via Cloudflare tunnel: ringcatch.io, dashboard.ringcatch.io. Source is private (real client-facing business) - happy to walk through the architecture and code directly.

Infrastructure & Homelab

Hardware

debianbox was archbox (Arch Linux) on the same hardware until wiped and reinstalled as Debian 13 on 2026-07-26 after its last Arch update broke reboot reliability. The cluster was originally deliberately three different distros (Debian, Arch, CentOS Stream) to keep manifests honest about distro-specific assumptions — since the rebuild it's Debian twice and CentOS once, so that property is weaker than originally designed. centosbook still doubles as a standing environment for engaging with the RHEL/CentOS Stream ecosystem.

Current Architecture (updated 2026-07-25)

                         INTERNET
                             |
                    [Cloudflare Tunnel]
                      (no open ports)
                             |
┌────────────────────────────────────────────────────────────────────┐
│  k3s CLUSTER  (LAN: 192.168.4.x, k3s v1.36)                       │
│                                                                     │
│  ┌────────────────────────────────────────────────────────────┐   │
│  │  MikePC  192.168.4.54  RTX 5060 Ti  (control plane + GPU) │   │
│  │                                                             │   │
│  │  namespace: ai                                              │   │
│  │  ├─ ollama [GPU]  gemma4:26b · gemma4:e4b · qwen3:30b     │   │
│  │  │               qwen2.5:7b · nomic-embed-text             │   │
│  │  ├─ pv-workbench      → http://pv.lan                      │   │
│  │  ├─ ams-intelligence  → http://ams.lan                     │   │
│  │  ├─ argus-bot         → Discord (Argus#1432)               │   │
│  │  └─ traefik ingress                                         │   │
│  └────────────────────────────────────────────────────────────┘   │
│                                                                     │
│  ┌────────────────────────────────────────────────────────────┐   │
│  │  debianbox  192.168.4.45  i3-4130T  (24/7 worker)          │   │
│  │                                                             │   │
│  │  namespace: agency  (21 of 24 services)                    │   │
│  │  ├─ orchestrator · outreach · scraper                      │   │
│  │  ├─ command · discord · billing · legal · marketing        │   │
│  │  ├─ support · success · bi · sales · cfo                   │   │
│  │  ├─ inbox · delivery · video · dashboard                   │   │
│  │  ├─ n8n · calcom · kokoro · voice                          │   │
│  │  └─ postgresql-16 (hostPath PVC)                           │   │
│  └────────────────────────────────────────────────────────────┘   │
│                                                                     │
│  ┌────────────────────────────────────────────────────────────┐   │
│  │  centosbook  192.168.4.33  CentOS Stream 10  (24/7, lid-closed) │
│  │  namespace: agency — landing + tunnel (no dependency on the │   │
│  │  shared debianbox-pinned agency-data-pvc; site survives a   │   │
│  │  debianbox outage)                                          │   │
│  └────────────────────────────────────────────────────────────┘   │
└────────────────────────────────────────────────────────────────────┘

LLM Routing (agency):
  Gemini 2.5 Flash → Ollama gemma4:26b (k3s GPU pod) → Groq llama-3.3-70b → Groq llama-3.1-8b

LLM Routing (pv-workbench):
  REASON_MODEL  gemma4:26b   signal detection, regulatory Q&A, MedDRA
  DRAFT_MODEL   gemma4:e4b   ICSR narratives, digests
  CHAT_MODEL    qwen2.5:7b   intent routing, Discord general chat

Where It's Going

Cluster is production-stable. Next phase is AWS hybrid: stateless public-facing services (landing, chatbot, webhook handlers) move to AWS free-tier EC2 while stateful workloads, LLM inference, and clinical AI stay on-prem. Tailscale bridges the two without open ports or VPC peering. Same Kubernetes manifests, new node pool.

homelab-infra

Full IaC for the cluster: k3s manifests for both namespaces (ai + agency), NVIDIA device plugin setup, Traefik ingress, GitLab CI/CD pipelines. Migrated 24 agency services from Podman quadlets to k3s on 2026-05-31. See k3s-homelab for the full day-one build documentation including lessons learned.

Dotfiles & Linux

dotfiles - Fish, Sway, Kitty, Starship, Waybar for the Rocky Linux daily driver. install-rocky.sh sets up a new machine in one command; legacy Debian/Hyprland config kept for history.

linux-notes - Running log of things worth remembering: hardware quirks, driver fixes, networking gotchas. RTX 5060 Ti Blackwell on Debian, Tailscale/nftables interaction on Arch, i915 firmware dimming, Chromebook UEFI.

Stack

Python · FastAPI · LangChain · ChromaDB · Ollama · Streamlit Kubernetes / k3s · Traefik · GitLab CI/CD · Podman (image builds) Cloudflare Tunnel · Tailscale · nftables · CrowdSec · AdGuard Home PostgreSQL · n8n · Terraform · AWS EC2 Debian 13 · Rocky Linux · CentOS Stream · Sway · Fish