UC AI Stack Schematic

UMBRACORE · APPLIED CYBER-PHYSICAL LAB

UC AI Stack Schematic

Current-state system drawing and specifications for the self-hosted AI stack on the lab LAN: two GPU inference nodes, a five-container service tier on Proxmox, the private YaCy index, a Proxmox Backup Server off the storage array, a Tailscale gateway for remote access, a planned MailHub for local email, the UmbraCore website (staging in the lab, production on a Linode), and no cloud model route connected.

Dwg no.

UC-AI-001

Revision

M · 2026-10-18

Baseline

baseline-20261004

Owner

Dennis Cary

Class

Public · scrubbed

System schematic SHEET 1

LAN request path Fallback only Private data / index Leaves the lab Local zone Restricted Cloud Parked / unassigned

Every request from a client, agent, or workflow enters through CT160 LiteLLM. Two paths leave the LAN: SearXNG public search queries, and the website push from ucweb to the Linode over SSH. The website is not part of the AI stack and holds no private data; it publishes only what is approved on staging. Gemini is not connected, so no model traffic reaches a cloud service. LiteLLM, n8n, CrewAI, Vane and YaCy all run from uc-ai-array iSCSI storage, so the array and its single 1 GbE link sit under the whole service tier. Backups leave the array for pbs-1, a separate box on the same LAN. Remote access comes in through Tailscale on dcmint, an encrypted tunnel that lands inside the LAN, so what a tailnet device can reach is set by the Tailscale ACLs, not the UDM Pro. MailHub (planned) will hold a local copy of personal mail. Its only outbound traffic is fetching from the mail providers, and its triage calls stay on Dexter through LiteLLM. Becky has no fallback and no cloud path.

Compute and storage nodes SHEET 2

Dexter

Hardware profile 2026-10-01

Role: Primary inference and RAG. Keeper node. All private data lives here.
Chassis: Micro (HK) Tech "AI Series" mini-PC · board Meigao F8BAC · BIOS AMI 1.04 (2025-04-11)
CPU: AMD Ryzen AI 9 HX 370 (Zen 5) · 12C / 24T · 5.16 GHz max · 24 MB L3 · XDNA NPU (unused)
Memory: 96 GB DDR5-5600 SO-DIMM (2 × 48 GB Micron) · both slots used
dGPU: NVIDIA RTX 5070 · GB205 · 12 GB GDDR7 · 250 W limit · compute 12.0
GPU link: PCIe x4 @ 2.5 GT/s (capable x16 Gen4/5). Runs through a riser.
iGPU: Radeon 890M (drives the HDMI console)

Storage: Crucial P3 2 TB NVMe · 1% wear · root LV 100 GB, ~74% used · ~1.7 TB unallocated in VG
Network: 2 × RTL8125 2.5 GbE in bond0 (both linked at 1 GbE) · Wi-Fi 7 down
OS: Ubuntu 26.04.1 LTS · kernel 7.0.0-34 · Docker 29.8.0
GPU stack: nvidia-driver-595-open 595.91.07 · CUDA 13.2 · LM Studio CUDA runtime (Vulkan retired)
LM Studio: Headless llmster under user dc · :1234 OpenAI-compatible · openai/gpt-oss-20b only · 64K context
AnythingLLM: Docker (docker run, not compose) · v1.16.1 · :3001 · multi-user mode · storage /opt/anythingllm/storage · NAS share \\nas\AI-Share
Workspaces: dexter (Chief-of-Staff) · becky (restricted)

VRAM fit: gpt-oss-20b weights are roughly 12–13 GB in MXFP4, so on a 12 GB card part of the model or KV cache likely sits in system RAM at 64K context. Check the live split with lms ps and nvidia-smi before deciding on a larger card.

Gadget

Model spec 2026-10-08

Role: Current: failover for dexter and stack, direct gadget route. Mostly idle.
Chassis: Minisforum AtomMan G1 Pro · 3.8 L vertical mini-tower · 215 × 57 × 315 mm · 3.81 kg
CPU: AMD Ryzen 9 8945HX (Dragon Range, Zen 4) · 16C / 32T · 5.4 GHz max · 80 MB total cache · iGPU drives one HDMI
dGPU: NVIDIA RTX 5060 · 8 GB GDDR7 · 145 W TDP
Memory: 32 GB, one SO-DIMM (slot P0 Channel B, module rated 5600 MT/s) · second slot empty · single-channel · board max 2 × 48 GB DDR5-5200
Storage: 2 × M.2 2230/2280 PCIe 4.0 NVMe, 4 TB per slot max · Installed drives not captured
Power: Internal 350 W PSU · 245 W sustained CPU+GPU (100 + 145) · Office / Gaming / Beast profile switch

Cooling: Glacier gen-3: dual 8015 fans, five copper heat pipes, rated 300 W
Network: 5 GbE RJ45 (Realtek RTL8126) · Wi-Fi 7 · BT 5.4 · Switch-port speed not captured
Display out: 2 × HDMI 2.1 (1 iGPU, 1 dGPU) · 2 × DP 2.1b · 1 × DP 1.4a · up to four 4K
USB: Front: 1 × USB-A 10 Gb, 1 × USB-C data · Rear: 2 × USB-A 10 Gb, 1 × USB-C data
OS: Windows 11 Pro · bare metal (Linux conversion under consideration)
Serving: AnythingLLM :3001 · workspace assistant-chats · Qwen3-VL 8B Instruct (vision) · up from 4B on 2026-10-10
Planned: Blueprint (Oct 7) makes Gadget "The Worker": n8n, Perplexica, vision/OCR via LM Studio. Not implemented.

Expansion limit: the 350 W internal supply is already budgeted at 245 W sustained for CPU and GPU, and the chassis has no open slot, so a second GPU in this box is not an option. VRAM growth means an external GPU or a different node. Run Gadget in Beast mode for inference and check its thermals under a long load. VRAM fit: Qwen3-VL 8B at Q4 is about 5–6 GB of weights plus the vision projector, which leaves roughly 2 GB for KV cache on the 8 GB card. Keep context modest, and confirm nothing spills to the single-channel system RAM.

Proxmox cluster · 4 × LattePanda Sigma

Captured 2026-10-08

Nodes: uc-ai-ve-1 + 3 identical nodes · Other node names not captured
System: LattePanda Sigma SBC · BIOS 5.27
CPU / node: Intel Core i5-1340P (Raptor Lake-P) · 4P + 8E = 12C / 16T · Iris Xe iGPU
RAM / node: 32 GB LPDDR5, soldered (8 × 4 GB channels) · not upgradeable
Disks / node: WD_BLACK SN770 500 GB NVMe · iSCSI LUNs 2 TB + 1 TB from uc-ai-array (shown on uc-ai-ve-1)
Network / node: nic0 1 GbE up · nic1 down · Wi-Fi down
Cluster total: 48 cores / 64 threads · 128 GB RAM · ~2 TB local NVMe
Platform: Proxmox VE 9.2.21 · kernel 7.0.14-23-pve
Stack guests: CT160–164 · unprivileged Ubuntu 24.04 · nesting=1,keyctl=1 · onboot · Current node not confirmed (uc-ai-ve-1 reports 0 CTs)
Guest storage: CT160 LiteLLM, CT162 n8n, CT163 CrewAI, CT164 Vane and YaCy run on iSCSI LUNs from uc-ai-array Stated · CT161 AnythingLLM Not captured
Restore pts: baseline-20261004 (CT160–164) · yacy-20261004 (CT164) · pbs-1 backups Not captured

uc-ai-array

Captured 2026-10-09

Role: iSCSI block storage for the Proxmox cluster. Holds the disks for LiteLLM (CT160), n8n (CT162), CrewAI (CT163), Vane (CT164) and YaCy. If it stops, the gateway stops, and every route stops with it.
Model: QNAP TS-932PX · 9 bays (4 × 3.5" + 5 × 2.5")
Firmware: QTS 5.2.10
CPU: Annapurna Labs Alpine AL324 · 4 × ARM Cortex-A57 @ 1.7 GHz
RAM: 16 GB (15.8 GiB visible, 13.1 GiB available) · the model's maximum
Network: eth0 LAN @ 1 GbE, all traffic (≈256 GiB in / 354 GiB out) · eth1 LAN @ 1 GbE, same subnet, nearly idle · eth2, eth3 no link (the 1000 they report is not a live link) · bond0–3 defined, unused
Data arrays: md2 RAID 5 · 4 × 12 TB HDD · 512K chunk · 32.7 TiB (36.0 TB) usable · [UUUU]
md3 RAID 1 · 2 × WD Red SA500 · 3.45 TiB (3.8 TB) usable · [UU]
Cache: md1 RAID 1 · 2 × Crucial BX500 · 1.63 TiB · [UU] · mirrored, which QTS requires for read-write cache
System: md9, md13 (QTS system, all 8 disks) · md256, md321, md322 (swap) · all healthy, no failed members
LUN map: Not captured which array backs the 2 TB and 1 TB LUNs

Bay	Drive	Size	Use
1–4	Seagate ST12000DM0007 (3.5" SATA HDD)	4 × 10.91 TB	Data
5	Empty	—	—
6–7	Crucial BX500 2 TB (2.5" SATA SSD)	2 × 1.82 TB	Cache
8–9	WD Red SA500 4 TB (2.5" SATA SSD)	2 × 3.64 TB	Data
Raw / usable		54.6 TB / 39.8 TB	usable = RAID 5 HDD + RAID 1 SSD

Both copper ports negotiated 1 GbE, so the switch side is gigabit. The TS-932PX's two 10GbE SFP+ ports are the likely eth2/eth3 and have no link, so all iSCSI rides one gigabit port. A RAID 5 rebuild after one 12 TB drive fails reads all three survivors end to end, about a day on this CPU, with no redundancy during that time.

pbs-1

Captured 2026-10-10

Role: Proxmox Backup Server. Backup target for the cluster, on a different box than uc-ai-array
Chassis: Model not captured
CPU: Intel Xeon E-2314 · 4C / 4T (no Hyper-Threading) · 2.8 GHz · 1 socket
Memory: 32 GB (31 GiB visible, 2.4 GiB used, 28 GiB available) · no swap
Network: nic0 LAN, same LAN as the stack · 29 RX / 7 TX drops · Link speed not captured
PBS: Not captured version, datastore name and path, backup jobs, prune and verify schedules, encryption

Disks: 4 × Seagate BarraCuda ST2000DM008 2 TB (3.5" desktop drive, SMR) · each split 1 GB EFI + ZFS member
Pool: rpool · RAIDZ1 (7.27 TB raw, ~5.2 TB usable) · 283 GB allocated (3%) · 6% frag · dedup off · ONLINE
Filesystem: rpool/ROOT/pbs-1 on / · 206 GB used of 5.2 TB · no separate datastore dataset in the listing, so backups likely share the root dataset
Boot: EFI partition on all four disks · boots from the same pool that holds the backups

The 5.2 TB of space is plenty for this stack. The drives are the weak point. ST2000DM008 is shingled (SMR), and SMR drives slow badly under sustained writes. A ZFS resilver after a disk failure can take days, and with RAIDZ1 the pool has no redundancy until it finishes. Plan to swap in CMR NAS drives (IronWolf or WD Red Plus) as budget allows. Until then, keep a copy of the most important backups somewhere else.

dcmint · Tailscale gateway

Captured 2026-10-10

Role: Tailscale gateway into the lab LAN. The remote-access path for tailnet devices
OS: Linux Mint desktop · Version not captured
CPU: Intel Core i7-11700 (Rocket Lake) · 8C / 16T · 2.5 GHz base
Memory: 16 GB (15 GiB visible, 13 GiB available) · 2 GB swap, unused
Storage: WD PC SN530 256 GB NVMe · ext4 / 17 GB used (8%) · 512 MB EFI

LAN: enp0s31f6 LAN (onboard Intel) · 0 errors · 79,149 RX dropped of 8.2 M packets (~1%) · Link speed not captured
Tailnet: tailscale0 tailnet address · MTU 1280 · ~1.3 GB in / 3.2 GB out
Tailscale: Not captured advertised routes, exit node, SNAT, key expiry, tailnet ACLs

Remote access to the lab rides Tailscale through this node; the tailnet policy decides which lab services a remote device can reach.

MailHub · local mail hub

Planned · not deployed

Purpose: Download and keep a local copy of mail from all POP3 / IMAP accounts so Outlook reads one place Stated
Functions: Folder rules · scanning for problems · access for the AI stack Stated
Deploy: Containers on Proxmox, built by a run-and-done script Stated
AI: Triage on Dexter through LiteLLM. Moving it to Gadget was considered and set aside for now Stated
Not captured: CTID, IP, node, disk location, account list, mail store size, component choices

Suggested parts: Fetch: getmail6 or fetchmail · Store: Dovecot IMAP with Sieve for rules · Scan: rspamd + ClamAV · Triage: small worker calling LiteLLM Proposal
Route: Use stack, not dexter. The dexter route passes through AnythingLLM, which drops JSON schema and adds vault RAG and the Chief-of-Staff persona to every email. stack is LM Studio on the same box with structured output intact. Give MailHub its own virtual key limited to stack
Auth: Gmail and Microsoft 365 no longer accept a plain password for IMAP / POP. Plan on app passwords (Gmail) or OAuth2 (Gmail, M365). Keep credentials in a root-only 600 secrets file inside the container
Disk: Mail store on local NVMe or iSCSI, not the AI-Share SMB path. Back it up to pbs-1 from day one

Overlap: Mailsort (Sheet 2, Parked) is a Gmail-only classifier with a morning digest. MailHub covers the same job for every account, so Mailsort can likely be retired or folded into MailHub's triage step once MailHub is up.

UmbraCore website · ucweb → Linode

Built 2026-10-18

Role: Public site www.umbracore.net: portfolio and project field notes. Staging in the lab, production on a Linode. Edits happen on staging only
Staging: LXC ucweb · Ubuntu 24.04 · nginx + PHP 8.3-FPM (pool user ucweb) + MariaDB · WordPress 7.1.3 · /var/www/umbracore · DB umbracore_wp · prefix wp_
Production: Linode umbracore-web · Ubuntu 24.04.5 · 3.8 GB RAM · 79 GB disk · Nginx :80/:443 → Apache 127.0.0.1:8080 (mpm_event, remoteip) → PHP 8.3-FPM → MariaDB · DB wp_umbracore · /var/www/umbracore
TLS: Let's Encrypt (webroot) for www + bare domain · auto-renew with nginx reload hook · issued 2026-10-18, expires 2027-01-16
Hardening: Host and cloud firewalls · fail2ban (sshd, nginx-limit-req) · wp-login 6 req/min · xmlrpc.php blocked · no PHP in uploads · HSTS + security headers · DISALLOW_FILE_EDIT · unattended upgrades

Push trigger: wp-admin 🚀 Push to Live (mu-plugin, staging only) writes /var/lib/umbracore-push/trigger → umbracore-push.path → umbracore-push.service · CLI umbracore-push.sh --dry-run | --push
Push steps: Preflight (siteurl, core versions) → gzip backup of live DB (keep 10) → export staging DB with wp search-replace to https://www.umbracore.net → rsync wp-content (excludes cache, push plugin, mcp-adapter) → import → flush → HTTP 200 health check, auto-restore DB on failure
Transport: SSH key deploy@the Linode · one ControlMaster session per push · deploy may run only wp as www-data · wp-content owned deploy:www-data, setgid

Staging is the source of truth: each push replaces the live database, so comments or edits made directly on the Linode are lost. Update WordPress core on both sides to the same version; plugins and themes on staging, then push.

Supporting systems

Stated

YaCy: :8090 · ~300 GB index · private SearXNG engine (local mode, enable_http: true) · disk on uc-ai-array iSCSI
NAS: QNAP LAN · AI-Share SMB for Dexter · UCArray2 CIFS · Same unit as uc-ai-array not confirmed
Edge: Ubiquiti UDM Pro: gateway, NAT and firewall for the lab · WAN1 gigabit fiber, WAN2 5G failover · Firmware, VLANs, rules not captured
Network: UniFi switching · Tailscale mesh, entering the LAN through dcmint (LAN)
Monitoring: Prometheus / Grafana deployed (no GPU or LiteLLM dashboards yet)

Parked and spare

Not in service

Hermes Agent: LXC hc, Ubuntu 24, web dashboard. Wired to Dexter LM Studio + AnythingLLM RAG. Parked Oct 3: contends with AnythingLLM for Dexter's single GPU.
Gemini route: Not connected to the stack. No live LiteLLM route, no key in use. Presidio config kept for when a cloud route returns.
Mailsort: Built Oct 6 (Gmail classifier + 06:30 digest via local LLM). Deployment unconfirmed

Service containers SHEET 3

CT	Host	Endpoint	Runtime	Function	Key paths	Disk	Status
160	litellm	:4000	Docker: LiteLLM main-stable, Postgres 16, Presidio ×2	Single model gateway, virtual keys, PII guardrail (idle: no cloud route)	`/opt/litellm/{compose.yaml,config.yaml,.env}`	iSCSI	Verified
161	anythingllm	:3001	Docker: AnythingLLM, password set	Front door. Holds Dexter and Becky persona prompts	`/opt/anythingllm/{server.env,storage/}`	Not captured	Verified
162	n8n	:5678	Docker: n8n, SQLite	Orchestrator: research, morning brief, report outbox, crew relay	`/opt/n8n/{compose.yaml,data/,workflows/}`	iSCSI	Verified
163	crewai	:8000	Python venv, systemd `crewai-api`	Multi-agent tasks behind `/webhook/crew`	`/opt/crewai/{app.py,.env}`	iSCSI	Verified
164	perplexica	:3000	Docker: Vane (ex-Perplexica) + SearXNG	Cited web research; YaCy + public engines	`/opt/perplexica/{compose.yaml,searxng-settings.yml}`	iSCSI	Verified

Model routing · LiteLLM SHEET 4

Route	Backend	Upstream URL	Fallback	Allowed keys	Use
dexter	Dexter AnythingLLM ws `dexter`	:3001/api/v1/openai	Gadget `assistant-chats`	Client key	Default chat, RAG over vault
becky	Dexter AnythingLLM ws `becky`	:3001/api/v1/openai	None · fails closed	Client key	Restricted, manual chat only
stack	LM Studio direct, `openai/gpt-oss-20b`	:1234/v1	Gadget `assistant-chats`	All	Machine calls, @agent, JSON schema, tools
gadget	Gadget AnythingLLM	:3001/api/v1/openai	—	All	Direct vision model (Qwen3-VL 8B)
mailhub key	MailHub triage worker → `stack`	via :4000	None · triage waits	New `mailhub` key, `stack` only	Planned Email classification and summaries
gemini	Not connected No live route. Nothing in the stack reaches a cloud model.				—

Settings: turn_off_message_logging: true · store_prompts_in_spend_logs: false · drop_params: true with model_info on stack so JSON schema survives
Presidio: No route to guard MASK: person, email, phone, IP, driver license, medical license · BLOCK: SSN, ITIN, passport, bank, card, IBAN · tested fail-closed
Why two Dexter paths: AnythingLLM's OpenAI endpoint drops JSON-schema and tool options, so anything structured goes straight to LM Studio via stack.

Workspaces, agents and workflows SHEET 5

Personas

Where	Name	Chat	@agent
CT161 client	Dexter · Chief of Staff	dexter	stack
CT161 client	Becky · restricted	becky	stack
Dexter node	dexter, becky (RAG stores)	gpt-oss-20b
Gadget node	assistant-chats	Qwen3-VL 8B
Stack-wide	Confidant / thinking partner	Honest pushback, plain prose

Personas live in the client because AnythingLLM's API replaces the upstream workspace prompt with the caller's. Upstream RAG still applies.

Automation

Agent skills: uc-web-research → n8n /webhook/research · uc-crew-task → n8n /webhook/crew
n8n flows: research-library v2 · morning-brief v2 · report-outbox (→ UC Report Inbox in Obsidian)
CrewAI: FastAPI app behind systemd, model stack
Agent Flow Lab: Four teaching flows installed in AnythingLLM
MailHub: Planned fetch → rules → scan → triage via LiteLLM stack
Gaps: @agent skills not tested end to end from the chat box

Data boundaries SHEET 6

Restricted

Becky workspace on Dexter only
Manual chat only. No agents, n8n, CrewAI or web paths
No fallback, no cloud route

Local · private

Vault, RAG, YaCy, all inference on Dexter and Gadget
LiteLLM stores usage only, not content
Work and patient data excluded from this stack entirely
MailHub mail store and triage (planned): personal accounts only, never leaves the LAN

Egress

SearXNG public engines (queries)
Tailscale: encrypted remote access in through dcmint. An inbound path, not data egress
MailHub (planned): fetches from mail providers. Credentials go out, mail comes in
Website push: SSH from ucweb to the Linode. Carries only published site content
Gemini: not connected. No model traffic leaves the lab
Claude and other cloud AIs: no PII

Retired components SHEET 7

Ollama + Open WebUI: Replaced by LM Studio + AnythingLLM. "Gadget" persona and Hindsight memory layer went with it.
mem0 / Qdrant: Dropped after contamination and schema issues
WhiteRabbitNeo: Removed from Dexter; gpt-oss-20b is the only model
l-llm (old router on gap-02): Superseded by CT160 at LAN
Vulkan runtime: Replaced by CUDA on Dexter
Qwen3-VL 4B: Replaced on Gadget by Qwen3-VL 8B Instruct, 2026-10-10