AI PILOT PROGRAM · INTERACTIVE PROJECT SHOWCASE

Autonomous Plant Operations — From Pilot to Enterprise

3,500 MTPD cryogenic ASU · 14 AI agents · 37 incidents autonomously handled · 2,036 knowledge documents
Airgas
An Air Liquide Company
INTERACTIVE SHOWCASE
Project Plan Overview
Autonomous ASU Operations — Pilot to Fleet
Scope: 35+ ASUs
Duration: 5 phases
Status: Phase 2 validated
1 Discovery Wk 1-4 DONE 2 Shadow Mode 14 days VALIDATED WE ARE HERE 3 Human-in-Loop 30 days REQUESTING 4 Limited Ops 60 days PLANNED 5 Fleet Scale-Out 4 plants TARGET GATE GATE GATE GATE Complete Current Planned Phase Gate (measurable exit criteria)
What We're Building
402 Plant Sensors Temperature, pressure, flow, vibration, purity Physics Engine (14 Models) MAC, Columns, PHX, Expanders, Tanks 14 AI Agents (LLM Swarm) Safety (seq) → Optimize (parallel) → Finance Coordinator: 98 approved / 32 rejected Knowledge Vault 2,036 docs RCA Engine Root cause + citations Operator Dashboard (12 Views) Production, Storage, Transit, Billing, RCA...
Phase 3 Ask — What Changes
Without AI (Today)
Incident response: minutes
Knowledge: in people's heads
New plant: months of training
Unplanned trip: $150k+
With AI (Phase 3+)
Incident response: < 5 seconds
Knowledge: 2,036 docs, searchable
New plant: config file + auto-learning
Damage avoided: $200k+ per event
Proven: How the System Handled Extreme Cold (-25°C)
0s Detect 402 sensors flag freeze 1s Analyze Physics models freeze risk 2s Act Load cut to 85% 3s Protect Hospital supply secured 4s Justify -$1k revenue +$200k saved 5s Recover Ramp back in 18h Result: 6 agents, 13 steps, fully autonomous, 0 safety events, $200k+ damage avoided Every step recorded in Knowledge Vault — next plant knows this playbook automatically
Key Workstreams
SAFETY & GUARDRAILS Kill switch, safety case, action boundaries HUMAN IN THE LOOP Operator training, feedback, trust NPS Phase 3 SECURITY & GOVERNANCE Threat model, pen test, MOC, ethics ROI & FINOPS Baseline lock, per-claim math, CFO signoff Ongoing FLEET SCALE-OUT 4 plants, cross-plant learning, config-driven Phase 5 BEYOND ASU HR, Marketing, Finance, Sales expansion Future
The Ask
1. Approve Phase 3 (HITL) at pilot site
2. Fund fleet architecture for 4-plant rollout
Full 25-section plan → Tab 5
25 plan sections · 19 RACI roles · 10 consultants · 5 phase gates with measurable exit criteria
Confidential

The Business Case for Autonomous Operations

ROI proof, competitive positioning, risk mitigation, and the cost of inaction

💰 Return on Investment

The pilot has demonstrated measurable value across three dimensions: avoided downtime losses, optimized load scheduling (off-peak fill vs. peak demand), and reduced operator response time from minutes to seconds.

Value LeverMechanismImpact
Incident ResponseAutonomous cascade (avg 10 agents in <5s)37 incidents handled
Load OptimizationOff-peak fill, demand-responsive scheduling98 load changes approved
Knowledge CaptureEvery decision traced & searchable2,036 docs generated
Alarm ManagementPattern recognition across 402 sensors280 alarm events traced
CFO signoff cycle: Monthly → Quarterly → Annual. Per-claim math requires Engineering + Finance co-signatures. Only A+B confidence claims count.

🏆 What We're Building

A reusable autonomous operations platform that scales across the fleet and captures institutional knowledge permanently.

CapabilityWhat It DoesStatus
Autonomous Operations14-agent swarm handles incidents end-to-endPilot validated
Knowledge ContinuityEvery decision captured, linked, searchable2,036 docs
Root Cause AnalysisAutomated incident investigation with citationsRunning
Fleet ArchitectureMulti-plant design, config-driven rolloutDesigned
Key differentiator: Full decision audit trail — every agent action has a traceable "why." Not just automation, but explainable automation.

⚠️ Risk Mitigation

The implementation plan is designed to minimize risk at every stage through progressive phase gates with measurable exit criteria.

RiskMitigation
Safety eventAdvisory-only (P1-P3), kill switch <30s, safety assurance case per IEC 61511
Operator resistanceNo-blame policy, union MoU, NPS tracking, change champion network
ROI shortfallExit write-down accounting, sunk-cost schedule signed at Phase 2 entry
Model driftEvalOps: 100+ golden scenarios, CI-triggered evals, results dashboard
Vendor lock-inOpen-weights models, OSS runtime, no proprietary dependencies

🚀 Scale Economics

The marginal cost of adding the next plant drops dramatically. The architecture, agents, knowledge vault, and training are already built.

ItemPlant #1Plant #2-4
Architecture & designFull cost$0 (reuse)
Agent developmentFull cost~10% (config only)
Knowledge vault2,036 docsCross-plant learning
Edge hardware1 server + GPU1 server + GPU each
Operator trainingFull program~30% (lessons learned)
FinOps projection: 3-year run-rate from pilot → 4-plant fleet. Per-plant/product/agent cost attribution with chargeback. Downside scenario modeled at 2× token costs.

How Autonomous Operations Work

A 3,500 MTPD cryogenic air separation plant, monitored by 402 sensors, operated by 14 AI agents, with every decision recorded in a living knowledge vault.

The Data-to-Decision Pipeline
🌡️
Sensors
402 ISA-tagged instruments monitor temperature, pressure, flow, vibration, purity across all equipment
402 live tags
⚙️
Physics Engine
14 equipment models (MAC, columns, PHX, expanders, tanks) with CoolProp thermodynamics
14 equipment models
🤖
Agent Swarm
10 core agents + 4 extended. Safety-first sequential checks, then parallel optimization
14 agents, 353 decisions
📚
Knowledge Vault
Every decision, alarm, and cascade step captured in interconnected Obsidian documents
2,036 documents
📊
Operator UI
12 operational lenses: production, storage, transit, delivery, billing, RCA, and more
12 lenses
Two Real Incidents — Autonomously Resolved
1

Extreme Cold Event (-25°C)

INJ-00001 · WEATHER
Injection EngineDetects -25°C. Flags cooling water freeze imminent.
MPC OptimizerCalculates optimal load cut — CW at -20°C, freeze risk critical.
Physics KernelReduces plant load 100% → 85%. GOX/GAN rates drop ~15%.
Dispatch AgentSuspends non-P1 deliveries. Pre-positions trucks at hospitals.
Economics AgentRevenue impact: -$1,087. Freeze damage avoided: $200k+.
Recovery PlannerSchedules ramp-back when temp returns above 0°C. ETA: 18h.
Outcome: Plant protected, P1 customers served, $200k+ damage avoided. 13 cascade steps, 6 agents, fully autonomous.
2

Compressor Vibration Spike

INJ-00002 · EQUIPMENT
Sensor AgentMAC vibration exceeds alarm threshold. MAC-TT-HI fires.
Physics KernelAssesses bearing stress, surge margin. Recommends 20% load cut.
Production AgentAdjusts column feed. Re-balances HP/LP reflux ratios.
Customer ImpactIdentifies at-risk deliveries. Flags 2 P2 customers for delay notice.
Loss LedgerEconomic impact: -$2,800/h. Trip cost if unaddressed: $150k+.
Recovery PlannerMaintenance window identified. Gradual ramp-back over 4h.
Outcome: Trip avoided, bearing damage prevented, maintenance scheduled proactively. 10 cascade steps, no unplanned downtime.

How the LLM Reasons Through Each Cascade

LLM + RAG + Root Cause Analysis

Agent Reasoning (LLM Cascade)

Each agent receives the plant state, sensor readings, and prior agents' outputs. The LLM reasons step-by-step, producing a structured decision with justification:

Physics Kernel — thinking
MAC vibration at 12.3 mm/s (alarm: 10). Bearing DE temp rising 2°C/hr. Surge margin at 15% — if vibration increases, we hit 10% margin and auto-trip fires. Reducing load by 20% brings vibration below alarm, opens surge margin to 28%, and buys 4h for maintenance assessment. GOX/GAN production drops proportionally but all P1 contracts remain fulfillable from tank inventory.
Economics Agent — thinking
Revenue loss at -20% load: $2,800/h for estimated 4h = $11,200. Cost of unplanned trip: ~$150,000 (emergency maintenance + 24h restart + lost production + customer penalties). Cost of load reduction is 7.5% of trip cost. Recommendation: Accept the load reduction.

Every agent's reasoning is captured in the knowledge vault. Future operators can search "compressor vibration" and see exactly why this decision was made.

Root Cause Analysis (RCA)

After the incident stabilizes, the RCA system performs a deep investigation using vector search (FAISS) across the knowledge vault and historical incidents:

Causal Chain:
Bearing wear Vibration increase MAC-TT-HI alarm Surge margin risk Load reduction Production impact
RCA Engine — analysis
Similar event: INJ-00017 (73.39h) and INJ-00034 (129.48h) both involved MAC compressor vibration. Pattern: vibration events recur every ~55h of high-load operation. Recommendation: Schedule predictive bearing inspection at 50h intervals. Estimated prevention cost: $2,400. Estimated avoided losses: $8,400+ per event.
Citations: INJ-00002 cascade trace, INJ-00017 cascade trace, INJ-00034 cascade trace, MAC equipment profile, Alarm/MAC-TT-HI history (26 activations)
25
Plan Sections
5
Phase Gates
19
RACI Roles
10
Consultants
37
Incidents Handled
353
Agent Decisions
100%
Drag to pan · Wheel to zoom · Click a section to jump

One Framework, Every Department

We built an autonomous agent swarm that runs a 3,500-ton/day cryogenic air separation plant. The same pattern — sensors, agents, knowledge vault, decision cascades — works for any domain where people make decisions under uncertainty. This is not a concept — it's running, validated, and audited.
10
Agents Proven
37
Incidents Handled
353
Cascade Decisions
2,036
Knowledge Docs
The Reusable Pattern
Domain
Sensors / Inputs
Simulation /
Model Layer
Agent
Swarm
Knowledge
Vault
Historian
DB
UI
Lenses
Replace the physics kernel with any domain model. The agent coordination, knowledge graph, incident learning, and cascade tracing work identically.
One architecture. Every department. Infinite institutional memory.
🏭
ASU Operations PROVEN
Air Separation Unit · 3,500 MTPD · Running Now
Agents
Production Agent Dispatch Agent Economics Agent MPC Optimizer Physics Kernel Recovery Planner Loss Ledger Customer Impact Tank Inventory Injection Engine
Example Cascade
Extreme cold (-25°C) detected Physics Kernel cuts load to 85% Production Agent flags tank drain risk Dispatch Agent suspends non-P1 deliveries Economics Agent costs the $1,087 revenue impact Recovery Planner schedules ramp-back in 18h
Knowledge Vault
2,036 documents · Equipment profiles, alarm definitions, 37 incident case files with full cascade traces, agent decision logs, audit reports, RCA investigations
Click to expand ↓
👥
HR & Workforce NEXT
Same agent pattern applied to people operations
Agents (mapped from ASU equivalents)
Recruiter Agent Onboarding Agent Compliance Agent Retention Agent Workforce Planner Training Coordinator
Example Cascade (like ASU incident response)
Certification expiring in 30 days Compliance Agent flags requirement Training Coordinator schedules session Retention Agent assesses flight risk Workforce Planner checks backup coverage Onboarding Agent prepares cross-training if needed
Knowledge Vault
Employee profiles ↔ Skills ↔ Certifications ↔ Training history ↔ Project assignments ↔ Team structures. Every compliance event traced like an ASU incident.
Click to expand ↓
📣
Marketing Automation NEXT
Campaign orchestration using the agent swarm
Agents (mapped from ASU equivalents)
Campaign Agent Content Agent Analytics Agent Budget Agent Audience Agent Channel Optimizer
Example Cascade
Campaign CTR drops below threshold Analytics Agent diagnoses drop pattern Content Agent proposes creative refresh Audience Agent segments underperformers Budget Agent reallocates spend Campaign Agent executes A/B test
Knowledge Vault
Campaigns ↔ Channels ↔ Audiences ↔ Content assets ↔ Performance metrics ↔ Competitive intel. Full history of what worked and why.
Click to expand ↓
💰
Finance & Controlling NEXT
Automated financial operations and audit
Agents (mapped from ASU equivalents)
AP Agent AR Agent Forecasting Agent Audit Agent Treasury Agent Compliance Agent
Example Cascade
Invoice variance > 5% detected AP Agent flags anomaly Audit Agent traces to purchase order Forecasting Agent adjusts cash flow model Treasury Agent repositions short-term instruments Compliance Agent logs for quarterly review
Knowledge Vault
GL accounts ↔ Cost centers ↔ Vendors ↔ Contracts ↔ Invoices ↔ Compliance rules. Every variance investigation becomes institutional memory.
Click to expand ↓
📝
Sales Proposal Automation NEXT
Multi-agent proposal generation & pricing
Agents (mapped from ASU equivalents)
Estimator Agent Scope Agent Risk Agent Pricing Agent Compliance Agent Proposal Coordinator
Example Cascade
New RFP received Scope Agent parses requirements Estimator Agent pulls similar past projects Risk Agent flags novel scope items Pricing Agent builds margin model Compliance Agent checks bonding/insurance
Why This Fits
Proposals are 200+ page documents with interdependent cost, schedule, and technical sections. The swarm pattern lets specialist agents work in parallel (structural estimating, MEP takeoff, schedule logic, commercial terms) then a coordinator synthesizes — exactly like the ASU agent cascade.
Knowledge Vault
Past proposals ↔ Win/loss analysis ↔ Scope templates ↔ Pricing models ↔ Client history ↔ Competitor intelligence. Every proposal outcome feeds back into the next one.
Click to expand ↓
Show / Hide
Total Documents2,036
Active Agents14
Incidents Tracked37
Cascade Steps353
Alarm Activations280
Coordinator Decisions130 (98 approved)
Click any node in the graph to see its details
Agents
Equipment
Incidents
Alarms
Bus Events
Cascade Steps
Index
WHY THIS MATTERS
Every decision, every alarm, every incident response — captured, connected, and searchable. This is how institutional knowledge scales.

Scales With Your Fleet

Each new plant gets its own equipment, sensor, and alarm subgraph. Cross-plant incidents automatically link to multiple plant nodes. 4 plants × 402 sensors = 1,600+ ISA tags, all traced and connected. Adding a plant takes a config file — the vault grows itself.

Incident Pattern Recognition

Every incident creates a full cascade trace — which agents responded, in what order, with what economic impact. When the same event hits plant #3 two years later, the vault surfaces exactly how plant #1 handled it. 37 incidents resolved, 102 scenarios cataloged, 353 decision steps recorded.

Knowledge That Doesn't Retire

When an experienced operator leaves, their knowledge walks out the door. This vault captures every decision, every alarm response pattern, every edge case. New operators see the reasoning behind past actions, not just the actions themselves. It's a living training manual that grows with every shift.

Audit & Compliance Ready

Auditors can trace any alarm → to the incident that caused it → to the agents that responded → to the economic outcome. Full chain of custody, always current, never in someone's email. Red-team reports, verification reports, and RCA investigations all link back to their source equipment.