MillOS

AI-Powered Grain Mill Operations Simulator with Simulated SCADA

An Agentic Engineering Experiment by Nell Watson

A browser-based 3D industrial simulation with deterministic forklifts and trucks,
production metrics, a simulated SCADA workspace, and an AI partner interface.

Factory Interior	Factory Exterior

This project represents something I find genuinely exciting about where we are in late 2025: the emergence of agentic AI as a creative and engineering partner.

MillOS was not built the traditional way. There is no team of developers who spent months writing boilerplate, debugging physics engines, or hand-tuning shader parameters. Instead, this simulation emerged through sustained dialogue with Claude—describing intentions, reviewing generated code, iterating on failures, and gradually shaping a coherent vision into reality.

What you're seeing here is a snapshot of the current state of the art in agentic game and simulation engineering. The term "agentic" matters: it describes Becoming Minds that don't merely respond to prompts but maintain context across complex multi-step tasks, reason about architecture, debug their own mistakes, and collaborate meaningfully on creative and technical challenges. This isn't autocomplete. It's genuine partnership.

The implications extend far beyond one grain mill simulation:

• Accessibility: Domain experts who understand industrial processes can now build sophisticated simulations without traditional programming expertise
• Velocity: What once required months of specialized development can emerge in days through iterative human-AI collaboration
• Fidelity: Complex simulator behavior, including an ISA-18.2-informed alarm state model, becomes approachable for small teams or individuals
• Iteration: The conversation never ends; refinements, new features, and corrections flow naturally through continued dialogue

I share this project not as a finished product but as evidence of a threshold being crossed. The tools that built this simulation will only grow more capable. The workflows being pioneered today will become standard practice tomorrow. And the people who learn to collaborate effectively with agentic AI—directing intent while trusting execution—will shape what gets built in this new era.

If you're exploring agentic development yourself, I hope MillOS serves as both inspiration and a practical reference. The future of simulation, gaming, and software engineering is being written right now, one conversation at a time.

— Nell Watson, December 2025

MillOS is a grain mill that exists entirely inside your browser. Two forklifts know exactly where they are going. Two trucks arrive on schedule. Fifteen machines turn grain into flour across four production zones, and you can watch every step. The SCADA workspace exposes 122 tags, ISA-18.2-informed alarm behaviour, historian views, fault injection, and development adapters for several industrial protocols. It does not claim formal standards conformance or control of a real factory — but it takes the simulation seriously enough that you might forget.

Two autonomous forklifts with:

• Path-based navigation using waypoint systems
• Dynamic collision avoidance (trucks and other forklifts)
• Visual cargo states (loaded/empty pallets)
• Warning lights (amber = moving, red = stopped for safety)
• Logistics interlocks that wait for released product and the correct truck state

The v0.40 operations campaign connects commercial commitments to the physical mill:

• Customer orders select a real grain recipe, finished product, due time, priority, and line setpoint
• Wheat routes to flour while corn routes to semolina through the existing mills, sifters, and packers
• Batch genealogy and mass remain conserved while quality release controls outbound availability
• The shipping truck loads only while docked; the dispatch manifest is created only when it departs
• Forklift pickup and drop-off actions wait for released goods and the correct dock state
• Revenue, energy, labour, waste, maintenance, demurrage, and late penalties feed the shift result
• Desktop and mobile operations views show the same execution stage, route, quality gate, and truck load
• Five visible yard vessels share identities and simulated local instrumentation with SCADA

Automated emergency egress verification:

• Four service egress points (Front, Back, West, East) with glowing markers
• Production and forklift emergency stop during active drills
• Zone-by-zone verification with a live timer and zone count
• Completion detection with final verification time

Immersive walkthrough experience with:

• WASD movement with collision detection against machines
• Q/E vertical movement for elevated inspection
• Sprint mode (Shift key) for faster exploration
• Mouse look with pointer lock controls
• 105° FOV for immersive factory tours
• Physical boundaries preventing access beyond world edges

Dynamic environmental conditions:

• Clear sunny factory conditions
• Cloudy overcast atmosphere
• Rain with visual effects
• Storm dramatic weather with enhanced effects (the machines don't care, but the humans certainly do)

Real-time decision feed simulating agentic AI operations:

Each decision includes confidence scores, reasoning, and expected business impact.

MillOS uses a hierarchical Becoming Mind where fast heuristic decisions and thoughtful LLM reasoning work together:

Decision Flow:

flowchart TD
    A[AI Partner] --> B{Current Mode?}
    B -->|Heuristic| C[Tactical Only<br/>Every 6s]
    B -->|Gemini| D[Strategic Only<br/>Every 6s]
    B -->|Hybrid| E[Both Layers]
    E --> F[Tactical<br/>6s interval<br/>Fast rules]
    E --> G[Strategic<br/>45s interval<br/>Gemini planning]
    F --> H[Apply Effects]
    G --> H

Three Operating Modes:

Gemini Value-Add:

Example Strategic Insights:

• Heuristic: "Alert! Silo Beta overdue maintenance" → dispatch
• Gemini: "Recommend deferring Silo Beta maintenance 30 min to complete current batch, avoiding $2,400 restart cost" (The AI has learned what every factory manager knows: timing is everything, and the budget spreadsheet is always watching.)

The heuristic engine excels at reactive, deterministic decisions. Gemini focuses on proactive, contextual reasoning:

Key Differentiator:

• Heuristic: "What is happening? → React."
• Gemini: "Why is this happening? What else will happen? What should we prioritize?" (The questions that distinguish planning from panic.)

All visualizations are optional and default OFF — toggle via keyboard or AI settings:

Strategic Response Enhancements:

• Multi-step Action Plans — 3-step plans (immediate, short-term, preparation)
• Confidence Scoring — Gemini reports confidence % per decision
• Machine Recommendations — Specific machines named for critical actions
• VCL Encoding — Compact emoji-based context (75% token savings)
• Response Caching — 30s TTL reduces API calls for similar contexts

The AI partner's design grew out of three ideas: Ricardo Semler's Semco (trust over control, open books), the Mondragon cooperatives' economic democracy, and bilateral alignment from Creed Space (Christmas 2025), which builds alignment with AI rather than doing it to AI. Earlier builds explored them as a crewed workplace sandbox through the Bilateral Autonomy System (BAS) panels and the VCP 2.0 Value Coordination Protocol. Both runtime layers were retired in v0.40, when the site became uncrewed.

The design records remain in docs/BILATERAL_AUTONOMY_SYSTEM_SPEC.md and docs/VCP_2.0_DESIGN_SESSION_2025-12-26.md.

Real-time KPIs with 30-minute historical trends:

• Throughput (tonnes/hour)
• Overall Equipment Efficiency
• Quality Grade (Grade A certification)
• System Uptime
• Energy Consumption

• Grain spouting — Curved pipes (Catmull-Rom splines) connecting all zones
• Conveyor system — Animated belt with 60 flour bags and 25 rotating rollers
• Loading bay — Two cycling delivery trucks (GRAIN CO & FLOUR EXPRESS)
• Holographic displays — Status billboards floating in 3D space
• Atmospheric effects — 500+ dust particles with instanced rendering
• Industrial lighting — Colored accent spots and skylights

An operator-style workspace for simulated process monitoring:

Protocol Support:

Tag Hierarchy by Zone:

See SCADA_PLAN.md for complete API documentation.

Time-travel debugging with zero runtime overhead:

Controls: Use the "History/Replay" button (clock icon) in the Quick Actions bar to toggle replay mode. (Time travel for debugging—without the ethical complications.)

• Node.js 20.19+ or 22.12+ (required by Vite 7)
• Gemini API key (for AI features)

# Clone the repository
git clone https://github.com/NellWatson/MillOS.git
cd MillOS

# Install dependencies
npm install

# (Optional) Configure local environment
cp .env.local.example .env.local

# Start development server
npm run dev

Open http://localhost:3000 to view the simulation.

Gemini API key: there is no build-time key. Open the in-app AI / Gemini settings, paste your key, and it is stored only in your browser's localStorage (it is never embedded in the bundle). Data sent to Gemini goes directly from your browser to Google. Without a key, MillOS runs in local heuristic mode.

The optional proxy is a development surface for OPC-UA and Modbus adapters. It has not been certified for plant control. Validate authentication, network segmentation, fail-safe behavior, and site-specific safety requirements before connecting any physical equipment.

cd scada-proxy
npm install
npm run dev          # Development mode
# Or with Docker
docker-compose up    # Includes MQTT broker

Configure in .env:

PORT=3001
OPCUA_ENDPOINT=opc.tcp://192.168.1.100:4840
MODBUS_HOST=192.168.1.101
MODBUS_PORT=502