Phase 1 · Progress updates

Phase 1 updates (10/12)

Rack rebuilds, a working Mini-to-Brume network proof of concept, and the next reliability tests.

Future plan

Finish the experiment rig, establish repeatable baselines, and test security before a limited student pilot.

%%{init: {"themeVariables": {"fontSize": "20px"}, "flowchart": {"nodeSpacing": 30, "rankSpacing": 45}}}%%
flowchart TD
  A["Finish Rack 2<br/>Boot · OS · GPUs"] --> B["Record a reproducible setup"]
  B --> C["Inference engineering<br/>Baselines and stress tests"]
  B --> D["Basic cyber testing<br/>Prompt injection and access controls"]
  C --> E["Compare models and serving methods"]
  D --> F["Test defenses and log failures"]
  E --> G["Review evidence"]
  F --> G
  G --> H["Limited student pilot<br/>Streaming API · keys · health checks"]

• Experiment rig.

  Rack 2 keeps disruptive experiments separate from Rack 1's shared inference service.

• Evidence before rollout.

  Follow Revollo's crawl, walk, run checkpoints; promote configurations only after performance and security review.

Rack 1 inference testing plan

Find the quality–latency–resource Pareto frontier, then validate the best configurations under concurrent use.

%%{init: {"themeVariables": {"fontSize": "20px"}, "flowchart": {"nodeSpacing": 30, "rankSpacing": 45}}}%%
flowchart TD
  A["Fixed coding workload"] --> B["Cold and warm baselines"]
  B --> C["Compare models<br/>Size · quantization · quality"]
  B --> D["Compare serving methods<br/>Speculative decoding · disaggregation"]
  D --> E["GPU placement and batching"]
  C --> F["Measure quality, speed and memory"]
  E --> F
  F --> G["Concurrent users and scheduling<br/>Queueing · fairness · stability"]
  G --> H["Load and recovery tests<br/>Sustained · burst · soak"]
  H --> I["Pareto frontier<br/>Choose primary and fallback"]

• Comparable results.

  Keep the workload fixed and repeat each configuration; raw throughput alone can hide poor quality or unfair scheduling.

• Hardware compatibility.

  Treat speculative decoding and disaggregation as experiments, contingent on runtime support and the GTX 1070/PCIe constraints.

What we have been working on

Rack 1 has been rebuilt into a cleaner, usable system, and Rack 2 reached BIOS for the first time.

%%{init: {"themeVariables": {"fontSize": "20px"}, "flowchart": {"nodeSpacing": 30, "rankSpacing": 45}}}%%
flowchart TD
  A["Rack 1: split-power experiment"] --> B["Full rewire and rebuild"]
  B --> C["Cleaner, usable rack"]
  C --> D["Next: reliability testing"]
  E["Rack 2: reworked with Mr. Mort"] --> F["First BIOS boot"]
  F --> G["Next: experiment-rig setup"]
  H["Mac mini network setup"] --> I["Brume laptop online<br/>Next: reboot, security and throughput"]

• Rack 1 rebuild.

  We replaced the split-power-supply experiment with a full rewire and rebuild; sustained-load and recovery testing comes next.

• Rack 2 milestone.

  With Mr. Mort's help, the reworked rack booted into BIOS for the first time, unblocking experiment-rig setup.

• Network setup on Mac mini.

  Built-in Wi-Fi and shared Ethernet now carry a laptop’s web traffic through the Brume 2. See the network diagrams below.

Mac Mini + Brume 2: laptop online

Proof of concept verified: a wired laptop loads Google through the Mini’s school Wi-Fi uplink.

flowchart TD
  A["School Wi-Fi<br/>PUSD-Onboard"] -->|"Built-in Wi-Fi · wlp2s0"| B["Ubuntu Mac Mini<br/>Shared mode: DHCP + forwarding + NAT"]
  B -->|"Ethernet · enp3s0f0<br/>10.42.0.1/24"| C["Brume 2 WAN<br/>Observed DHCP lease: 10.42.0.185"]
  C --> D["Brume private LAN<br/>192.168.8.x · second NAT"]
  D --> E["Wired laptop<br/>Web browsing VERIFIED"]
  E --> F["Next: reboot without USB dongle<br/>Persistence NOT YET VERIFIED"]
  classDef verified fill:#164b35,stroke:#56c997,color:#ffffff
  classDef pending fill:#343b48,stroke:#aab4c4,color:#ffffff
  class B,C,D,E verified
  class F pending

• Two roles.

  Mini = school-facing uplink. Brume 2 = private OCS LAN gateway. Addresses shown are observations, not the final address plan.

• Security gap.

  The working built-in Wi-Fi profile uses the unencrypted side of the school network. Firewall hardening and the EC2 tunnel are still pending.

What unblocked the network

Identify the adapter → load its driver → find a working connection → share Ethernet.

flowchart TD
  A["Adapter discovery"] --> B["0a5c:4500 = Bluetooth<br/>Not the Wi-Fi adapter"]
  A --> C["USB RT5370 · 148f:5370<br/>rt2800usb · OWE connection works"]
  A --> D["PCI BCM4360 · 14e4:43a0<br/>Found by Broadcom vendor ID"]
  D --> E["Installed broadcom-sta-dkms<br/>6.30.223.271 + headers + DKMS"]
  E --> F["wl loaded · wlp2s0 visible<br/>School SSIDs detected"]
  F --> G["Cloned / reused / pinned OWE profiles<br/>Connection attempts failed"]
  G --> H["Dedicated open profile<br/>Connected through built-in card"]
  H --> I["brume-link: shared Ethernet<br/>Brume lease + laptop web test passed"]
  classDef passed fill:#164b35,stroke:#56c997,color:#ffffff
  classDef blocked fill:#632f36,stroke:#e998a4,color:#ffffff
  class G blocked
  class F,H,I passed

• Observed workaround, not a finished uplink.

  The setup notes identify legacy wl / OWE incompatibility. The open profile is pinned to one access point; the USB dongle remains the working OWE fallback.

• HTTP works; ping failed.

  ICMP filtering is a hypothesis, not a confirmed diagnosis. Recheck DNS and HTTPS independently.

From working cable to dependable classroom network

Only the first milestone is complete. Each remaining gate needs recorded evidence.

flowchart TD
  A["DONE<br/>Mini → Brume → laptop online"] --> B["NEXT · Persist<br/>Reboot without dongle<br/>Autoconnect + DNS + HTTPS"]
  B --> C["MEASURE<br/>Built-in vs USB vs Rig 1 baseline<br/>Upload · download · latency"]
  C --> D["HARDEN<br/>Address plan + persistent forwarding<br/>Mini / Brume firewall + admin password"]
  D --> E["EXPAND LAN<br/>DHCP reservations + separate Wi-Fi AP<br/>Confirm ports / switch needs"]
  E --> F["CONNECT AWS<br/>EC2 ↔ Mini WireGuard<br/>Subnet routing + NGINX services"]
  F --> G["PROVE RECOVERY<br/>Wi-Fi loss → tunnel reconnect<br/>Driver repair + USB fallback runbook"]
  G --> H["THEN MIGRATE<br/>Move Rig 1 off its NetBird path"]
  classDef done fill:#164b35,stroke:#56c997,color:#ffffff
  classDef todo fill:#343b48,stroke:#aab4c4,color:#ffffff
  class A done
  class B,C,D,E,F,G,H todo

• Measure before committing.

  Run iperf3 toward EC2, compare both Wi-Fi adapters with the existing Rig 1 USB + NetBird path, and test available wired uplinks. No throughput result is available yet.

• Confirm the deployment path.

  Check school IT approval, Mini model year and Secure Boot status. Decide whether WireGuard terminates only on the Mini or also on the Brume.

Next architecture: one private OCS classroom LAN

Dashed links are planned. The Mini-to-Brume uplink is the demonstrated foundation.

flowchart TD
  A["AWS EC2<br/>NGINX + WireGuard"] -.->|"Planned encrypted tunnel"| B["Mac Mini<br/>School Wi-Fi uplink"]
  B -->|"Verified Ethernet link"| C["Brume 2<br/>Private LAN + DHCP + firewall"]
  C -.-> D["AI rigs<br/>Reserved LAN addresses"]
  C -.-> E["Jarvis 3000"]
  C -.-> F["Presence"]
  C -.-> G["Separate Wi-Fi access point<br/>IoT / classroom clients"]
  E -.-> H["Smart Classroom"]
  F -.-> H
  G -.-> H
  E -.->|"Future camera / vision workload"| D

• Brume is the wired gateway.

  Plan a separate access point for wireless clients; confirm LAN port capacity and whether a switch is needed. Choose IoT/client SSIDs or VLANs before rollout.

• Future experiment.

  Test routing Jarvis camera workloads to a stronger vision-language model on the rigs, potentially using dedicated Rig 2 GPUs. This is not implemented.

What still stands between the demo and deployment

Working Internet access is the milestone; resilience and security remain the next tests.

flowchart TD
  A["Open Wi-Fi + exposed admin risk"] --> A1["TLS / planned WireGuard<br/>Default-deny school-side ingress<br/>Strong Brume admin password"]
  B["Single pinned AP + one Mini<br/>Legacy DKMS driver"] --> B1["Test roaming and reboot<br/>Health checks + driver recovery<br/>Keep USB fallback"]
  C["Shared Wi-Fi + double NAT<br/>Mini Ethernet limited to 1 Gb/s"] --> C1["Benchmark simultaneous use<br/>Test wired uplink<br/>Validate routed access / discovery"]
  D["School policy + subnet collisions"] --> D1["Confirm with school IT<br/>Document non-overlapping subnets<br/>Reserve device addresses"]

• No production-readiness claim yet.

  Boot persistence, throughput, firewall rules, IoT Wi-Fi, EC2 routing, and recovery all need acceptance tests. Double NAT also needs validation for inbound services and mDNS discovery.

• Gateway follow-through.

  Replace the shared-mode proof of concept with reviewed persistent nftables/forwarding rules, test tunnel reconnection, and document wl recovery after kernel updates. A dedicated router may replace the Mini later.

Planning references

Plans draw on the project issues and Revollo's MVP checklist; test results are still pending.

• Issue #1 — Dual-rack hardware and software allocation
• Issue #4 — Inference bottlenecks and baseline metrics
• Issue #16 — Revollo's MVP checklist slide
• Issue #20 — Blueprint, model bake-off, stress tests, and pilot gates
• Issue #22 — Revollo meeting notes and weekly playbooks