Product APEX · Autonomous Plant Engineering Expert

APEX runs the plant.People approve.

APEX (Autonomous Plant Engineering Expert) is WI.Plat’s autonomous AI-agent platform for water facilities. APEX Operator runs the plant — planning pumps and valves 24 hours ahead against level, flow and tariff. APEX Engineer designs each site’s control from its own drawings and tags. Both run above the existing SCADA, inside the utility’s network.

Operator · autonomous operationEngineer · control designMPC optimisationAbove SCADA / PLCOn-premise sLLM
APEX Operator dashboard at a Korean water treatment plant: an isometric map of the plant's supply systems with reservoir levels, pump running status for each system, target-level achievement by reservoir, and daily electricity and carbon savings.
APEX Operator dashboard · water treatment plant, Korea. Eight supply systems, 34 pumps and 11 valves on one screen.
−15%
Energy charge · 3-day MPC simulation
34 pumps
Under control at one plant
14 plants
National water agency · awarded

Two modules, one platform

Engineer designs the control. Operator runs it.

One module produces the site-specific control design; the other executes it every day. What Operator learns flows back to Engineer for the next site.

Design module

APEX EngineerAI agent for control design

Reads the plant’s P&ID, HMI, tags, SCADA and documents; asks the operator what it cannot find; writes the autonomous-control design — state, action, constraints, objective — and generates the design package for people to review and approve.

  • P&ID · HMI · tag understanding
  • Constraint elicitation
  • Design package generation
Sites 1–2 people design, APEX learns · Sites 9–14 APEX designs, people approve

design → · ← operating data
Operations module

APEX Operatorautonomous plant & pump-station operation · ex-WAS-AI

Forecasts demand and reservoir levels, optimises the pump-and-valve schedule against the hourly tariff with model predictive control, recommends or — once approved — executes the setpoints, and verifies the saving against the baseline.

  • Demand & level forecasting
  • MPC optimisation
  • Recommend → approved control
  • Savings M&V
In service one Korean WTP, 8 systems · Awarded 14 plants · national water agency

Common agentic AI platformOn-premise sLLMLangGraph agent workflowKnowledge graph · Neo4jMultimodal RAG · Qdrant

The problem

Moving water is one of the largest electricity bills a utility pays.

Water and wastewater take about 4% of the world’s electricity, most of it in pumps. Which units run, at what speed, in which hour decides a large part of the operating budget.

4%

Of global electricity demand is used by water and wastewater services.

70–80%

Of a water utility’s electricity typically goes into pumping.

40%

Electricity can reach up to 40% of a water utility’s operating cost.

Operation depends on the operator on shift

Fixed scenarios run on personal experience. Patterns differ from person to person, and energy is wasted in the gap between them.

Tariff and demand are rarely weighed together

Deciding how much to pump in a cheap hour versus an expensive one, against tomorrow’s demand, is a calculation people cannot run every 15 minutes.

Pumps run outside their efficient combinations

The same water can be delivered by several pump combinations whose specific energy differs by a third — the choice is often habit.

Equipment faults go unheard

Legacy vibration sensors miss valve-seating faults and cavitation. The acoustic signature of a pump’s health is not used.

APEX Operator · how it runs the plant

Operator uses cheap and expensive hours strategically.

Every 15 minutes Operator reads the plant’s state, re-forecasts demand and levels, and re-optimises the next 24 hours — inside the constraints the utility has approved.

01

Collect

02

Forecast

03

Optimise

04

Recommend or execute

05

Verify

Fill in the cheap hours, draw down in the expensive ones

Reservoir target level · m · one system, one day · against the hourly tariff
Target levelOff-peak · KRW 82/kWhMid · KRW 87–106Peak · KRW 106

Tomorrow’s hourly target levels are set at 23:00; the pumps follow the curve. Tariff tiers: Korean industrial time-of-use, 2025.
Table view

The same water, moved like this

Shift in weekly pumped volume · m³ · one measured week, reference plant
Off-peak
+4,910
Mid
−3,118
Peak
−1,812
Same total volume — pumping moved from mid and peak hours into off-peak.

Same water, different bill

Specific energy by pump combination · kWh/m³ · one supply system

0.249 vs 0.327 kWh/m³ — a 31% spread for the same volume. Codes: running state of each of three pumps.
Table view

MPC operating structure

Current state → Plan → Verify → Execute → Compare with measurements · re-plan
Current state · operating conditionsSCADA level · flow · pressureAvailable pumps · valvesHourly electricity tariffOperating constraints
AI demand forecast

Clear-water tank and reservoir outflow, external customer take-off — 24 hours ahead.

AI candidate reductionauxiliary

Only verified pump combinations and operating ranges enter the search — a smaller candidate set, faster solve.

Optimisation solver · MIP
Objective
Minimise electricity cost or specific energy over the horizon
Decision
Target inflow per plant and reservoir · pump combination · VFD frequency · valve opening
Constraints
Clear-water and reservoir levels · pressure · equipment conditions

plannot met
EPANET hydraulic check

Supply flow, pressure and head computed for the plan; hydraulic feasibility verified, corrections returned to the solver.

verified · approved
SCADA execution

Target inflow, pump combination, frequency and valve opening applied through the existing control layer.

Measured comparison · updateForecast vs measured levelsSafety marginEquipment status changes→ periodic and event-based re-plan from the current state

Operator sits above the existing SCADA/PLC layer. Interlocks, trips and manual override remain the authority.

APEX Operator · MPC simulation case

Same three days, same water — 15% lower energy charge on paper.

One supply system, three days, re-planned by MPC every ten minutes against the operator’s actual record — same demand, same power model, same tariff. Simulation only; not yet a verified saving.

Three-day comparison · one supply system · 432 ten-minute steps
Indicator Operator · actual APEX MPC Difference
Electricity consumption 38,951.6 kWh 33,767.9 kWh −13.3%
TYPE1 energy charge KRW 4,569,264 KRW 3,880,370 −15.1% apparent
Pump running hours · 3 units 125.97 h 104.83 h −21.1 h
Pump starts · 3 units 6 21 +15
Level-bound violations · 10-min points 19 FAIL · operating bounds
Scope · domestic supply system
3
Pumps2 variable-speed + 1 fixed-speed
3
Storage nodesClear-water tank · two service reservoirs
27
Allowed control casesof 240 configured
13
Re-plans in 72 hscheduled and level-triggered

Service reservoir level — operator vs MPC

Level · m · hourly · 72 h · operating band 2.60–3.50 m
Operator · actualAPEX MPC · simulatedOperating band

Table view

Cumulative TYPE1 energy charge

KRW million · hourly · same fitted power model for both
Operator · actualAPEX MPC · simulated

Table view

Electricity-cost reduction · what has been sustained, simulated and expected
3.0%
Sustained · 2024 record · 8 systems
3.3%
Conservative scenario
15.1%
MPC simulation · 3 days · apparent
7–10%
Expected in operation · MPC engine
−3%
Electricity cost

Sustained in routine operation across 8 systems, 2024.

−20%
Pump alternation cycle

Automated alternation spreads the load across units.

−60%
Reservoir-level excursions

Tighter level control against the daily target curve.

APEX Engineer · how the next plant gets designed

An AI agent that designs APEX Operator for each site.

We taught the machine the process itself — how an expert reads a site, asks what is missing, and writes the control design.

What Engineer takes over
  1. Understand plant material

    P&ID, HMI screens, tag lists, SCADA configuration and operating documents.

  2. Ask what it does not know

    Asks the operator for the missing constraint or rule instead of inventing one.

  3. Design the autonomous control

    State, action space, hard constraints and objective for each hydraulic system.

  4. Generate the design package

    The Operator configuration for the site — people only review and approve.

The human role shifts as sites accumulate · share of design done by APEX
SITES 1–2

People design · APEX learns

The expert’s own design process becomes training data.

SITES 3–8

APEX drafts · people review

Repeatable parts are reused; only site-specific constraints are asked.

SITES 9–14

APEX designs · people approve

The design package is generated; human effort moves to review and approval.

“At the first plant APEX learns from people; at the last plant APEX Engineer performs most of the design.”This is not a systems-integration model that staffs every site. The 14 plants are where APEX learns, is verified, and becomes standard.

Common platform · security & autonomy

Inside the closed network — an on-premise agent.

Both modules run on one agentic AI platform — language model, agent workflow, knowledge graph and document store, all on-premise. In public procurement, that security is the barrier to entry.

On-premise sLLM

Answers operator questions, explains why a schedule was chosen and writes reports — without external cloud inference.

Closed network

Agent workflow

Classifies the request, calls the right skills and models, queries the data stores and composes the answer.

LangGraph

Knowledge graph

Four knowledge maps of the plant — facility, network, AI model and multimodal — so reasoning follows causality, not just correlation.

Neo4j

Multimodal RAG

Manuals, drawings and maintenance photos retrieved live to ground every answer in the site’s own documents.

Qdrant

IoT sensors

Acoustic and pressure sensing on the network and in the plant, feeding the same platform as SCADA.

Sensor network

AI models

Supervised and unsupervised models for forecasting and anomaly detection; MPC for optimal control.

CNN · AE · DQN · MPC

Data stays with the utility

Tags, flows, levels, pressures and energy remain on the utility’s own infrastructure.

Standard interfaces

OPC UA/DA, Modbus TCP or API to the existing PLC/SCADA. A pump disabled for maintenance drops out of the optimisation.

Explains itself

The agent supports the operator with reasons — forecast, tariff, constraint. It does not issue commands on its own.

From proof to contract

Verified at one plant, selected for fourteen.

2024

Reference WTP · 8 systems

APEX Operator (then WAS-AI) deployed and verified in daily operation.

2024–25

Design basis established

Transition from Q-table reinforcement learning to model predictive control.

2025

Wastewater plant proof of concept

Extension from water supply to wastewater treatment.

APEX process view for a wastewater treatment plant: power, unit power, temperature, flow, dissolved oxygen and total nitrogen at the top, with inlet gate, screens, pump tank, blowers, anaerobic, anoxic and aerobic tanks and membranes along the treatment train.
Wastewater, the same way. The same forecast-and-optimise loop applied to blowers and the biological process train.

Deployment

From diagnosis to control, one gate at a time.

Each stage needs the utility’s written acceptance; a utility can stay in advisory mode as long as it wants.

STAGE 01

Site assessment & data audit

READ-ONLY
STAGE 02

Baseline definition

READ-ONLY
STAGE 03

Design & integration

READ-ONLY
STAGE 04

Shadow-mode validation

SHADOW
STAGE 05

Advisory operation

ADVISORY
STAGE 06

Approved closed loop & M&V

CLOSED LOOP

Adoption

Three ways to bring APEX in.

From a conventional implementation to a contract paid only from verified savings.

System build

APEX, installed

Performance-based

Paid from verified savings

Integration

Water AI-platform integration

Find out what your pumps could save.

Twelve months of electricity bills and a tag list are enough for a preliminary energy assessment.

Contact WI.Plat