updown move enter open esc close matheesan.ca / index
    node moon-01 gpu rtx a2000 cameras 09 storage 22tb status nominal
    Homelab topology: a Proxmox host with an RTX A2000 running Ollama, Immich ML, Frigate for nine cameras and Whisper; a 10G switch, 22 TB ZFS storage, and services including Immich, Jellyfin, Authentik and observability.
    moon-01 / topology · not to scale

    The model is the easy part.

    Production systems taught me that intelligence is the demo; reliability is the job. The loop below is the part I actually build: models wrapped in constraints, tools, traces and evals that make failure visible before a caller finds it. The tools change; the loop stays.

    agent_loop.yaml production
    01 / input

    listen

    telephony · ASR · barge-in

    audio in
    02 / policy

    decide

    LLM intent + deterministic guards

    state machine
    03 / tools

    act

    APIs · idempotency · auth

    side effects
    04 / oracle

    verify

    evals · traces · must / must-not

    evidence out
    input: a real call loop: verify feeds the next release output: action + trace

    models

    • LLM tool-calling agents
    • prompt & state design
    • model upgrades as releases
    • ASR / TTS pipelines

    routing

    • provider fallback
    • cost / latency routing
    • retries & limits
    • per-tenant policy

    knowledge

    • ingestion & chunking
    • embeddings
    • vector search
    • citations

    evals

    • scenario harnesses
    • must / must-not assertions
    • silent-green corpus
    • replay & trace analysis

    agentic tooling

    • MCP servers
    • Claude Code skills
    • code-enforced guardrails
    • diff & blast-radius review

    infra

    • GCP Cloud Run
    • Terraform
    • BigQuery
    • GPU passthrough at home

    // ai adoption

    • 2023 GitHub Copilot — team talks
    • 2024 Cursor — engineering offsite + workshop
    • 2026 Claude Code — memory patterns

    “One of our most AI-forward engineers.” — manager’s farewell

    eval_console historical corpus · replay
    $ ./run-evals --suite silent-green
    6/16 field failures encoded 5 rules accepted 0 silent
    model_card.md local
    model
    matheesan-1
    context
    7 years AI/ML + production systems
    strengths
    evals · debugging · AI-native delivery
    limits
    will rewrite your test suite; says “it depends” when it does
    status
    building at WayTrade

    currently: marketplace platforms at WayTrade · local inference on a 12 GB GPU · MCP servers · evals for AI-authored config

    “Put deterministic logic where correctness matters. Put LLMs where judgment helps.”

    — the rule I keep relearning

    A production lab, under the moon.

    The homelab is where I test the ML I can’t put in production: local inference on a 12GB GPU, computer vision on nine cameras, and enough VLANs to make a network engineer smile. It is part curiosity, part skill-building, part just loving that it works.

    nvidia-smi passthrough ok
    GPUNVIDIA RTX A2000
    VRAM12 GB
    LLMqwen3.5:9b · qwen3:8b-16k
    VISIONYOLOv9m · LPR · face
    [gpu]

    Local inference

    RTX A2000 12GB passed through to the GPU VM. Ollama serves qwen3.5:9b, qwen3:8b-16k and nomic-embed-text; Whisper runs speech-to-text on CUDA; Open WebUI is the front door.

    RTX A2000OllamaWhisperCUDAOpen WebUI
    [cv]

    Computer vision

    Frigate NVR with YOLOv9m on nine cameras, license-plate recognition on the driveway, and face recognition — all on the isolated camera VLAN.

    FrigateYOLOv9mLPRReolinkTapo
    [ml]

    Photo intelligence

    Immich with CUDA-accelerated face and object recognition. Thumbnails on SSD, cache on a dedicated disk, originals on the NAS over NFS.

    ImmichCUDAPostgresValkey
    [px]

    Proxmox hypervisor

    AMD Ryzen 5 2600X, 62GB RAM. Multiple VMs with VLAN isolation and PCIe passthrough. Dual network: 1Gbps plus a 2.5Gbps direct link to the NAS.

    ProxmoxKVMZFSVLANPCIe
    [nas]

    22TB storage

    TerraMaster + OpenMediaVault. Separate volumes for media, personal files, VM backups and ISOs, exported over NFS with jumbo frames on the 2.5Gbps link.

    OMVNFS22TB2.5Gbps
    [net]

    Network & identity

    Five VLANs, GL.iNet Flint 3, managed switches, Tailscale subnet routing and split DNS. Authentik SSO, Vaultwarden, AdGuard and CrowdSec keep the front door boring.

    VLANTailscaleAuthentikAdGuardCrowdSec
    domain *.matheesan.ca — wildcard SSL via Let’s Encrypt + Route 53
    remote Tailscale VPN with subnet routing + split DNS
    models Ollama — qwen3.5:9b · qwen3:8b-16k · nomic-embed-text
    speech Whisper STT on CUDA · Frigate LPR + face recognition
    monitoring Prometheus · Grafana · Alertmanager · Uptime Kuma · Scrutiny
    secrets Vaultwarden self-hosted + Bitwarden CLI automation

    the lab is invite-only · authentik sso · access →

    The lab is invite-only.

    I run Jellyfin, Immich and a few other services for people I actually know. If you’d like access, reach out and I’ll set you up. One login, everything in one place.

    by invitation

    Request access

    Jellyfin, Immich & more — accounts are set up personally.

    jellyfin immich nextcloud drop box authentik sso

    secured via Authentik SSO — you’ll need credentials from me directly.