Avery Hauser · projects in progress

Most of what’s here runs on hardware I can unplug.

I’m a computer science student and vice president of the OHIO Cybersecurity Club. This is the working notebook for what I’m building: a local LLM rig on a Proxmox box, a privacy tool that handles my data-broker opt-outs, and the segmented lab that keeps it all apart.

Now building: Local inference on an RX580, via Vulkan
Always on: nightshade, a Raspberry Pi 5
Waiting on: A GPU decision

Solanaceae Atropa belladonna L. Coll. night-shade.cc · September 2026 Each fruit is a project. Hover to name it.

Projects

What’s running, what’s being built, and what’s waiting on parts. Open a card’s field notes for the details.

5 machines · 3 zones · 0 port forwards

The lab

Private services live on the tailnet. The club gets its own subnet, reachable only through Cloudflare Access. The malware bench never touches a network at all.

Inventory
Host	Hardware	Role	Zone
t14	ThinkPad T14 Gen 3 · i7-1270P · 32 GB · 1 TB	Daily driver; isolated, on-demand practice VMs	Tailnet
nightshade	Raspberry Pi 5 · 8 GB · 64 GB SD	Ubuntu Server; Tailscale exit node; OptOut, Actual Budget	Tailnet
pve	Ryzen 7 7700 · 16 GB · 2 TB SSD · RX580 8 GB	Proxmox VE 9; llama.cpp + dsh in a Debian 12 LXC	Tailnet, own subnet
t5610	Dell Precision T5610 · CPU upgrade pending	Club server behind Cloudflare Access	Club subnet (planned)
t470s	ThinkPad T470s	Malware analysis	Air gap (planned)

Open decision · undecided

The next GPU

The RX580 gets models running. It won’t get me to bigger models or training. The shortlist is two used cards, and each one gives up something different.

	In the box nowRadeon RX580 8 GB	UsedTesla V100 32 GB	UsedRTX 3090 24 GB
Memory	8 GB GDDR5	32 GB HBM2	24 GB GDDR6X
Bandwidth	256 GB/s	900 GB/s	936 GB/s
Architecture	Polaris	Volta	Ampere
Software	llama.cpp on Vulkan	Stuck on CUDA 12.x — CUDA 13 dropped Volta; no bf16	Current CUDA, bf16
Power & cooling	185 W, own fans	250 W, passive — needs forced airflow	350 W, big 3-slot cooler
The case for it	Already running	Biggest memory pool; room to add a second card later	Newer stack, easier to live with, less memory

Also on the table: two V100s now, then a newer-generation card once GPU and RAM prices come back down.

Updated September 2026

Now & next

On the bench

local-llmGetting llama.cpp’s Vulkan backend stable on the RX580 inside a Proxmox LXC.
harnessWiring DeepSeek Harness to the local model.
optoutRunning OptOut in review-only mode on the Pi, rootless and tailnet-only.
clubKeeping ohio-cyber.us current for the club.

Up next

gpuChoose between a V100 32 GB and an RTX 3090.
t5610Upgrade the CPUs and stand up the club subnet behind Cloudflare Access.
routerAdd a second router so the network can be split properly.
t470sTake the T470s offline for good and set it up for malware analysis.