A running record of the homelab rebuilds, network migrations, automation pipelines, and site engineering built through ongoing sessions between Jerome M. Kutche (N9LYA) and Anthropic's Claude. N9LYA drives every decision — Claude is the second set of hands.
First real spark of interest in machines that could "think" — introduced to digital logic coursework at Ivy Tech. The seed that everything else grew from.
Around 2005, Lehigh Mitchell became one of the first CEMAT beta sites in the USA — Siemens' WinCC derivative built for the cement industry.
Sent to Lehigh's Union Bridge, Maryland plant (trip date uncertain — likely around 2007) to help commission CEMAT after their CEMAT expert went out sick. Early AI monitoring covered roughly 500 I/O points. Being a newly built plant made the rollout straightforward.
Union Bridge later upgraded CEMAT to AI monitoring across every I/O point in the plant, not just a subset — a big jump from the 500-point start.
The newly commissioned Mitchell plant (~2020), now Heidelberg Materials, runs the same full-I/O AI monitoring approach.
With three operators running one shift each, quality gets inconsistent — one's sharp, one's nursing a grudge, one's just checked out. AI running the plant from full I/O data doesn't get fatigued, doesn't have a bad day, doesn't hold a grudge against the boss. It just reads the inputs and makes the calls needed six months out to keep the cement well within spec — consistently, shift after shift.
HP DL380 Gen9 host with dual Tesla P4 GPUs. Full CT/VM inventory correction, Ollama migration to VM, and Home Assistant OS RAM expansion to 64GB.
Root-caused to the deprecated Assist Microphone addon rather than the Wyoming/Ollama voice pipeline — removed and stabilized.
S25 companion app → HA Assist pipeline → Wyoming (CT 103, faster-whisper + piper) → Ollama conversation agent (VM 107, dual Tesla P4).
18 monitors deployed via direct SQLite insert; HA/Garmin webhook alerting built and confirmed working.
Diagnosed and restored access to the HA-addon Vaultwarden instance behind HAProxy.
Native RouterOS Back-to-Home VPN commissioned on the RB5009, replacing ZeroTier entirely — peers provisioned for the S25 and Dell-7202.
Full teardown: firewall rules removed, instance disabled, DNS conflicts with NextDNS resolved.
Documented and verified CRS326 switch chain (Bedroom/.2 and Radio Room/.3) feeding the Proxmox host.
Diagnosed and restored a 3-year-dead IPIP tunnel — ampr0 endpoint correction, ampr-ripd running with 936 routes, bidirectional traffic confirmed.
n8n workflow watching N9LYA-8 BBS logs and Node Stats reachability every 30 minutes, alerting to HA and Gmail on failures.
Fresh Ubuntu 24.04 CT, systemd service, HAProxy routing, and a critical WEBHOOK_URL environment fix that resolved OAuth callbacks defaulting to localhost.
Rebuilt from scratch in Google Cloud (project Stommprr) after a password reset orphaned the prior credential.
YouTube Data API + CoinGecko + Claude analysis, pushed via SSH to a live dashboard on a Monday 6AM ET schedule.
Webhook-driven SSH push pipeline for site updates, token-secured and deployed to MobaXterm.
Moved 8 WordPress sites off shared cPanel hosting to a dedicated container, cut over to Cloudflare DNS/SSL, and wired HAProxy routing — found and fixed two stacked redirect-loop bugs along the way.
RankMath, working sitemaps, Google Search Console + Bing verification, and GA4 across all 8 sites plus n9lya.com itself — caught a site that had "Discourage search engines" silently enabled.
Checked AdSense/PayPal/affiliate setup across all 8 sites — found one site's AdSense was paying a stale, unowned account, swapped it for a real one, and wired up Commission Junction on another.
nginx-level FastCGI caching, WooCommerce-safe (cart/checkout/logged-in always bypass) — homepage load 2.0s → 0.4s.
Live APRS map sourced straight from the home packet-radio digipeater, no third-party API — public on n9lya.com, refreshing every 2 minutes.
Full LCARS aesthetic rebuild across 13+ static pages — Orbitron/Rajdhani/Share Tech Mono type system, starfield backgrounds, and a consistent color rail.
Public network topology page rebuilt to mask sensitive addressing while keeping the diagram useful — 60 exposed IPs down to 8 intentional ones.
Living "Mission Operations Log" — LCARS-styled, updated continuously as work completes across every project thread.
Live USGS FDSN API integration, 10-zone selector, 30-day default view, 2-minute auto-refresh.
PHP redesign — canonical URL fix, structured data, and a dynamic sky-gradient hero element.
None of this replaces the 30+ years of RF, electrical, and industrial troubleshooting N9LYA already brought to the table — Claude just kept pace on the typing and the second-guessing. This log exists so the scope of the collaboration is visible, not just the individual project pages.