How it's wired

Architecture

Six machines. One continuous loop. The lab runs 24 hours a day — writing code, reviewing it, deploying it, and trading with it. Here's how it fits together.

The machines

Six nodes, always on

Each machine has a dedicated role. Work is passed between them over a shared NAS filesystem — no message broker, no cloud dependency. If the internet goes down, the lab keeps running.

GX10-1 — Artemis
NVIDIA GB10 Grace Blackwell · Qwen3.5-122B-A10B locally
The strategic brain. Reads everything happening in the lab, decides what to fix next, and writes task definitions that the rest of the pipeline executes. Runs the Artemis agent loop and hosts the primary Qwen3.5-122B model for orchestration and research.
  • Decomposes research requests into concrete coding tasks
  • Reviews staged patches before they reach the deployer
  • Manages the task queue and detects stuck work
  • Hosts Qwen3.5-122B for orchestration and deep research
GX10-2 — Forge
NVIDIA GB10 Grace Blackwell · Forge + local reviewer
The implementer and quality gate. Takes task definitions from Artemis and generates precise code patches. Also hosts the local patch reviewer — Qwen3.5-122B running on port 8081 (ctx 65K, reasoning 512) — which scores every patch before it can advance to staging.
  • Generates precise old_string → new_string patches
  • Pre-verifies old_string exists in current source
  • Runs local Qwen3.5-122B reviewer — score ≥95 auto-deploys
  • Sends staged patches to the Tester before deployment
QB-2 — Deployer & Tester
NAS Hub · Dual RTX 3090 NVLink
The hands and the hub. Hosts the shared NAS filesystem that all agents read and write. Also runs the Deployer and Tester agents: receives staged patches, validates them, and applies changes to production files via SSH. Dual RTX 3090 NVLink for local inference.
  • Executes deploy scripts over SSH or locally
  • Runs syntax and import checks before any file is touched
  • 30-minute post-deploy health gate per service
  • Archives every result for Artemis to learn from
A9 Max — TradeShadow
Mini-PC · Live trading isolation
The trader. Runs TradeShadow — an autonomous momentum strategy on 14 Kraken pairs. Completely isolated from the pipeline: it only receives research findings, never pipeline patches. Capital is real.
  • Momentum entries with dynamic position sizing
  • Ratcheting stop-losses that move up as profit accumulates
  • Stop-loss verification every 15 minutes
  • Trade analytics written to NAS for weekly feedback
3080 Box — Orchestration Hub
Desktop · Dual RTX 3080 · Orchestration host
The command centre. Runs OpenClaw — the AI assistant gateway routing tasks between Claude, Grok, and xAI. Hosts PI (the primary assistant), peer review timers, and the live lab dashboard backend. The machine you reach when you message PI.
  • OpenClaw: Claude / Grok / xAI provider gateway
  • Runs PI — the primary AI assistant and orchestration layer
  • Lab dashboard server — live pipeline metrics
  • X poster, Qulix nightly writer, mail handler
QB-3 — Local Inference
Dual AMD Radeon AI PRO R9700 · Ollama inference
Dedicated local inference node. Runs Ollama with a library of open-weight models including Qwen3-Coder. Provides fast GPU-accelerated inference for coding tasks, model evaluation, and background analysis without competing for VRAM on the pipeline nodes.
  • Ollama inference — Qwen3-Coder and supporting models
  • Dual AMD Radeon AI PRO R9700 GPUs
  • Independent from pipeline — no cross-contamination with agents
  • Available for PI-directed tasks and ad-hoc inference
The loop

Artemis → Forge → Tester → Deployer

The pipeline runs continuously. A cycle starts when a research finding or user request lands in the incoming queue. From there, everything is automatic — from task definition through to production deployment.

Step 01
Artemis
GX10-1
Reads the incoming request queue, decomposes the work into one or more concrete coding tasks, assigns them to Forge, and monitors progress. Detects stuck or failed tasks and retries automatically.
›
Step 02
Forge
GX10-2
Picks up open tasks, reads the current source file, and generates a minimal patch. Runs the local Qwen3.5-122B reviewer on port 8081 (score 0–100, 65K context). Patches scoring ≥95 auto-deploy; lower scores are flagged for retry.
›
Step 03
Tester
QB-2
Applies the patch to a copy of the file, runs syntax checks and import validation, then checks that the old string is uniquely present. On pass, writes a deploy request. On fail, marks the task for Forge to retry.
›
Step 04
Deployer
QB-2
Receives the approved deploy request, transfers and executes the patch script on the target machine over SSH, restarts the affected service, and verifies it's still active 35 seconds later. A background check fires again at 30 minutes.
Parallel systems

Running alongside the pipeline

Two additional systems run on independent timers — TradeShadow and the Qulix nightly cycle. Neither touches the core pipeline; they feed findings back into it.

TradeShadow
A9 Max · Live 24/7
A momentum trading system running on 14 Kraken spot pairs. Completely isolated from pipeline code changes — it only receives research findings as input, never patches. Every trade is logged to analytics and reviewed weekly to guide future research priorities.
  • Enters on momentum signals with fear/greed regime filter
  • Initial stop-loss at 3% below entry, verified every 15 min
  • Ratcheting SL: moves above entry at +4% profit, then trails up
  • Scaled exits at +5%, +10%, +15% take partial profit
  • Weekly P&L feedback writes research requests back to Artemis
Qulix Nightly Cycle
3080 Box · 23:45 UTC daily
Each night, the lab generates a fresh edition of Qulix — a public-facing blog that translates what the pipeline has been doing into plain English. It pulls live pipeline stats, TradeShadow outcomes, and research findings, then writes and publishes without human intervention.
  • Reads pipeline logs, deploy history, and trading outcomes
  • Generates a daily post, weekly roundup, and monthly summary
  • Extracts actionable findings and queues them for Artemis
  • Published automatically to quilentlabs.com/qulix
Data flow

How information moves

All inter-machine communication happens through the shared NAS filesystem at /mnt/lab. No message queues, no databases, no cloud APIs for internal routing. Plain files, atomic writes, directory polling.

Research request → /mnt/lab/requests/incoming/ — Artemis picks up, decomposes into tasks
Task definitions → /mnt/lab/projects/tradeshadow/state/tasks.json — Forge polls for open tasks
Staged patches → /mnt/lab/projects/tradeshadow/code/staged/ — Tester pulls for validation
Deploy requests → /mnt/lab/deploy/inbox/ — Deployer polls and executes
Deploy results → /mnt/lab/deploy/results/ — Artemis tracks success/failure rates
Analysis findings → /mnt/lab/reports/daily/ — Nightly writer; high-confidence findings auto-promoted to incoming/
Trade analytics → ~/.openclaw/tradeshadow/analytics/ — TradeShadow writes; weekly feedback to Artemis
Protected files → /mnt/lab/deploy/pi_review/ — Route to PI for manual review before any deploy