Eddy: Your AI Operations Partner

The Problem with "AI for Operations"

You've seen the demos. Some vendor shows up with an "AI solution" for industrial operations, and it falls into one of three predictable traps—usually all three simultaneously.

First, there are the generic chatbots that don't understand your specific facility. Ask "What's my DO in Basin 2?" and they pull a SCADA tag. That's it. No investigation into why DO might be concerning, no reasoning about what it means in context, no memory of similar situations. When you face a complex operational decision—the kind that actually matters—they're useless.

Then there are the black box recommendations. The system tells you to "Reduce waste rate to 225 GPM" with zero explanation. Why? Based on what data? What happens if you don't? What if conditions change? Operators don't trust what they can't understand, and they shouldn't. These systems ask for blind faith in critical operations.

Worst of all are the one-shot answers with no follow-through. The system makes a recommendation, doesn't verify what actually happened, doesn't learn whether it worked, and repeats the same mistakes next time. There's no accountability, no learning loop, no improvement over time.

Eddy: The AI Agent That Actually Helps

Eddy isn't a friendly chatbot. He's an AI agent built specifically for water operations—designed to investigate problems, recommend actions with full transparency, verify outcomes, and continuously improve.

How Eddy Works

1. Goes "In Flow" When Problems Arise

When the Rules Engine detects an issue, it creates a Wave—a tracked investigation from trigger to resolution. Eddy immediately goes "In Flow." He pulls the Codex Protocol specific to this problem—an investigation plan built from your operators' expertise—and breaks the complex issue into manageable steps. He deploys Tools from the Toolkit—calculations, models, queries—systematically working through the diagnostic process. Every finding gets documented. Every decision gets recorded. And the plan adapts based on what he discovers along the way, just like an experienced operator would adjust their approach based on what the data reveals.

Example Wave:

WAVE #1251: AvN Controller Optimization
Trigger: Control authority at 62% (threshold: 80%)
         Temperature forecast: -8°F over 3 SRTs

Eddy's Investigation Plan:
├─ Step 1: Baseline assessment ✓
├─ Step 2: Waste rate optimization (running...)
├─ Step 3: Swing zone testing (if needed)
└─ Step 4: Generate recommendation

Eddy deployed:
- Data Query Tool (pulled last 30 days trends)
- SRT Calculator (tested -25%, -50%, -90% scenarios)
- SUMO Model (simulated zone configurations)

Finding: Combined approach needed
├─ Waste rate reduction alone: 68% confidence
├─ Zone switching alone: 71% confidence
└─ Combined approach: 87% confidence

Based on 23 similar historical events with 89% success rate.

2. Recommends with Full Transparency

Eddy never acts autonomously. He investigates, analyzes, and recommends—then waits for operator approval. This isn't a limitation; it's fundamental to the design. Critical infrastructure can't run on autopilot.

Every recommendation comes with complete transparency: a confidence level based on data quality and historical success rates, a reasoning trace showing exactly how he reached this conclusion, the alternatives he considered and why they were rejected, expected outcomes with specific timeframes, and historical context from similar events your facility has experienced. You're never asked to trust a black box. You're given all the information needed to make an informed decision.

The Three Modes of Eddy

Mode 1: Waves (Structured Investigations)

Waves are triggered automatically by the Rules Engine when something needs attention. This is Eddy at his most systematic—following Codex Protocols with investigation plans, deploying tools in sequence, establishing human collaboration checkpoints, and setting up outcome monitoring schedules. Everything gets documented: every step logged, every tool deployment recorded, every operator decision tracked, long-term outcomes monitored over days or weeks. When you look back at a Wave six months later, you see exactly what happened, why decisions were made, and whether they worked.

Mode 2: Ripples (Conversational Assistance)

Ripples are operator-initiated—quick questions and ad-hoc analysis when you need an answer now. It's the conversational mode, but backed by real intelligence.

Operator: "What's my current F/M ratio?"
Eddy: "F/M is 0.28 lb BOD/lb MLSS·day (calculated 2 min ago)"

Operator: "What would happen if we reduce polymer dose by 20%?"
Eddy: [runs calculation using Flux Engine]
      "Belt press capture would drop to ~89%,
       increasing return load by ~450 lb TSS/day.
       Based on current operating conditions, not recommended."

Operator: "Show me DO trends in Basin 2 for the last week"
Eddy: [displays chart with annotations]
      "Average 1.8 mg/L, trending up. Zone 2a switched
       to aerobic 3 days ago per Wave #1251."

Mode 3: Background Monitoring

Even when you're not actively using Eddy, he's working. In the background, he continuously monitors data quality—checking sensor health and calibration status. He watches process trends, noting when things approach thresholds or exhibit unusual patterns. He verifies Wave outcomes, tracking whether implemented changes are producing expected results over time. And he looks for learning opportunities, extracting patterns that should update Codex templates to make future investigations smarter. When he sees something important, he creates a Wave or notifies operators—but never, ever acts without approval.

What Makes Eddy Different

Built on Plan-and-Execute Methodology

Eddy doesn't wing it. Every investigation follows a rigorous, structured approach. He pulls the relevant Codex Protocol for the situation and breaks it down into steps with clear objectives. He deploys Tools to gather data and run analyses. He incorporates operator observations and local knowledge—you're part of the process, not an afterthought. He generates recommendations with confidence levels and complete reasoning. He requires human approval before any action—no exceptions. After implementation, he verifies what happened through Wave Ledger's three-source reconciliation. He tracks long-term outcomes over days or weeks, not just immediate results. And he extracts learnings to improve future investigations, making the system continuously smarter about your facility.

Leverages Your Facility-Specific Intelligence

Eddy gets smarter about your plant specifically. Codex templates encode your senior operators' diagnostic methods—their experience becomes his knowledge base. Historical Wave data teaches what actually works at your facility, not theoretical best practices from a textbook. Tool success patterns reveal which calculations are most predictive for your process conditions. Seasonal adjustments get captured from your operational history, so the system understands how your plant behaves differently in summer versus winter. Anomaly baselines are learned from your normal conditions, not generic industry averages.

He's not generic. He's yours.

Transparent and Explainable

Every Eddy recommendation can be traced back to its sources. You can see exactly what data was used, what Tools were deployed, what calculations were run, what historical patterns informed the analysis, and why this approach was chosen over alternatives. Click through the reasoning trace and you'll find real logic, not AI hand-waving. No black boxes. No hallucinations. No unexplainable authority.

Never Acts Autonomously

This deserves emphasis because it's so fundamental to the design. Eddy NEVER changes setpoints, starts or stops equipment, adjusts chemical dosing, or takes any control actions. He investigates. He recommends. He waits for approval. He tracks outcomes. That's it. Human-in-the-loop isn't optional or a feature you can toggle—it's the foundational architecture of the system.

How Eddy Learns

Reality Verification

After you approve a recommendation and take action, Wave Ledger doesn't just assume everything went according to plan. It verifies what actually happened through three-source reconciliation: your explicit confirmation of what you did, automated logbook parsing that extracts actions from shift notes, and change detection algorithms analyzing the data to confirm when parameters actually shifted. All three sources must agree to establish confidence. If they don't align, the system flags it for human review rather than making assumptions.

No assumptions. Only verified reality.

Long-Term Outcome Tracking

Success isn't determined by implementation—it's determined by whether it actually worked. And "worked" doesn't mean just the immediate result. Wave Ledger monitors outcomes over time—Day 1, 3, 7, 14, 21 checkpoints—with automated metric evaluation at each stage. It analyzes trends, matches against success criteria, and builds a complete picture of whether the intervention produced lasting improvements or just temporary fixes.

Wave #1251 - Day 14:
├─ Control authority: 84% (target: >80%) ✓
├─ MLSS: 3,900 mg/L (range: 3,600-4,100) ✓
├─ Stability: Maintained for 14 days ✓
└─ Outcome: SUCCESS

Pattern extracted and fed back to Codex.
Next time this happens, Eddy is smarter.

Continuous Improvement

Successful patterns update Codex templates:

• "Combined waste rate + zone switching works better for cold weather events"
• "This waste rate scenario gives more predictive results"
• "Confidence should be adjusted based on sensor calibration status"

Every Wave makes Eddy smarter about your facility.

The Partnership Model

Eddy isn't replacing operators. He's amplifying their expertise:

• Operator plant knowledge + Systematic investigation = Faster root cause
• Process intuition + Scenario analysis = Confident decisions
• Historical memory + Data-driven patterns = Optimized strategies
• Final authority + Transparent reasoning = Trusted intelligence

Not a Chatbot. A Partner.

Eddy represents a fundamentally different approach to AI for water operations:

• ✓ Systematic investigation, not conversational hand-waving
• ✓ Verified outcomes, not one-shot recommendations
• ✓ Continuous learning, not static rules
• ✓ Full transparency, not black box magic
• ✓ Human partnership, not autonomous control
• ✓ Plant-specific intelligence, not generic AI

Tools that think. Interfaces that disappear. Operations that flow.