Skills and Systems

Local model serving, RAG, agent orchestration, routing, and evaluation. LoRA and QLoRA on image models.

• 32B diffusion model trained as a QLoRA on a 24 GB consumer GPU under Windows.
• Orchestrator grounds every answer in the lab's own documentation, with citations enforced through tool calls.
• Three-tier local model roster routed by task; ARM cluster runs small LLMs and VLMs as workers.
• Anything the AI changes about itself is promoted by a human only.

Evidence: Home AI (HAI), the orchestrator · 32B QLoRA · Gauntlet (model tournament) · Ensemble experiment · Native Intelligence

Field-service operations for a regional team of about thirty technicians.

• Scheduling, coverage, escalations, service levels, logistics, field resources.
• Hiring, coaching, performance management.
• Translated field problems into process and tooling changes: requirements gathered from technicians, escalated with vendors, verified in use.

Evidence: Doctrine · HAI Orchestrator

Professional repair and QC background, applied to whole systems.

• Root-caused a driver-level GPU stall to the on-GPU firmware scheduler after vendor tooling showed nothing.
• AI agents in the lab are held to a written failure taxonomy, each class tied to a dated incident.
• 688-cell model-by-hardware tournament, 512 cells run and 436 past the quality bar, positive controls on every batch. Result: three ARM worker tiers, one broadly deployable vision model plus a quality specialist on the largest device class.
• 1,172-arm ensemble experiment varying team shape from one to ten workers, model held constant. Execution nearly complete; the rest and the appraisal wait behind other work.

Evidence: AMD GPU Wedge · Gauntlet · Ensemble experiment · Doctrine

Windows Server administration, Hyper-V, containers, Linux when needed. Home Active Directory from 2023 to 2025.

• Media-stack VM and home-automation VM as separate Hyper-V guests.
• About thirty containers under Docker Compose; every config file in version control.
• Canary-first deploys: nothing reaches live until the test gate passes and a human promotes it. Automatic rollback on failure.

Evidence: Current Lab · HAI Orchestrator

Power, watchdogs, and guards that fail closed, across five hosts.

• Traced lab-wide reboots past the software checklist to utility flicker at the wall. Fixed with three UPS units, dependency-ordered shutdown, and live energy telemetry.
• Watchdog rule: a dead reporter is never treated as a dead worker.
• Rescue daemon for GPU stalls; order-taker releases its lane claims by every exit path.

Evidence: Doctrine · HAI Orchestrator · AMD GPU Wedge · Current Lab

Storage tiering and workload placement, measured in time-to-complete under real load.

• 46 TB spinning for the library, all-SSD working tier on the training host, a 12 TB docked drive for second copies of what deserves one.
• Ingest staged to SSD before HDD; separating ingest from archive cleared a throughput bottleneck.
• Fit a 32B model onto a 24 GB card through VRAM residency and memory management.
• Scheduler places work by declared capability, memory headroom, and energy price.

Evidence: Current Lab · 32B QLoRA · HAI Orchestrator

Routing, DNS, firewalls, Wi-Fi, remote access, multi-site.

• Cloud front door: droplet, reverse proxy, automatic Let's Encrypt, private overlay to home. Zero inbound ports at the house.
• Two DNS resolvers, synced nightly: LAN macvlan and droplet secondary.
• Subnet router for remote access; per-capability forward-auth at the edge.
• Wired USB-Ethernet segment for the ARM cluster's head node.
• Two physical sites run as one system over a private overlay network; the media library is served from whichever has the better uplink.

Evidence: Current Lab · Friend Onboarding · Fellowship (the ARM cluster)

Scripting, APIs, forms and data movement, operational handoffs.

• Infrastructure as configuration in git; scheduled tasks with restart policies and live-decoy verification.
• PowerShell and Bash tooling across the fleet, written with AI agents under my review.
• Lab work runs as product orders: inputs, method, deliverables, cleanup.
• Designed a Raspberry Pi Zero USB-HID automation for device intake: wipe, re-enroll, test.
• Designed a Forms and Power Automate inspection workflow (design stage; not deployed).

Evidence: Doctrine · HAI Orchestrator · Current Lab

Component-level CPU, GPU, memory, storage, and platform knowledge from years of builds, from salvaged parts to eight-GPU rigs with ECC memory. Professional repair and QC.

• Five x86 hosts sized to distinct roles. 24 GB RX 7900 XTX for training; RTX 3070 and 3070 Ti for the 8-9B tier and failover.
• Both 3070-class cards root-caused to an unstable factory boost profile. Clock-locked at 1500 MHz, they pass hour-long soaks.
• Power-policy and PCIe-ASPM fix found on one box, carried to every peer.
• Eleven-node cluster of low-power ARM devices, including a 32-bit device on a custom native build.

Evidence: Current Lab · GPU boost-profile investigation · Fellowship · 32B QLoRA

Photography and editing since 2021; video encoding and transcoding against format, codec, and bandwidth limits.

• Self-hosted media library for friends behind capability-tiered access, with self-service requests. Served off-site for upload bandwidth until fiber arrived at home.
• Photo-editing export pipeline that produced the adapter-training dataset.
• Audio transcription on the ARM cluster: a 29m54s episode in 14m13s on one node, 2.1× realtime, under a 35 W rated peak per node.
• Camera detection runs on the cluster when armed; devices hardware-decode 3072×1728 HEVC.

Evidence: Media Stack · Cameras · Fellowship · 32B QLoRA