Taking cost out of the tonnes you already make.

Your volume is set by the market.

An industrial plant at work

Three numbers your plant is already measured on.

Variance drives the other two.

01 · Cost per tonne

The same tonnes, cheaper

Fuel, power and freight for each unit of output, at the output you already sell. kcal/kg, kcal/kWh, rupees per tonne-kilometre.

02 · Variance

The same tonnes, every day

How far the plant moves from shift to shift.

Every swing is paid for twice: once as waste, once as the margin operators hold against the next one.

03 · Cash

The same tonnes, paid sooner

Money sitting in silos, spares stores, idle trucks and bills that were never reconciled.

Reduce the swing, then move the target.

In a cement kiln, free lime in the clinker is known an hour or more late.

A plant engineer and a data engineer working on one model

The model is built where they overlap.

The plant engineer knows which transmitters drift, why mill 3 always runs hot, and what the night shift does that nobody wrote down. They cannot hold four hundred tags and three years of history in mind.

The data engineer finds real patterns across those tags, but cannot tell a failing sensor from a real excursion. So every site team pairs both with your shift in-charge, who owns the decision.

Above your control system, never inside it.

The first question every plant asks is whether it can trip the unit.

01

Clamps in your control system

Configured in your DCS logic.

A mill bias stays within 5 to 10% of its own demand; an oxygen trim stays above an absolute floor.

02

Rate limits

No step changes.

Setpoints move slowly, and bad or frozen inputs freeze the bias.

03

A heartbeat

Our server updates a counter every second.

No update for 3 seconds and every bias ramps back to zero, your loops take over, and the operator is alarmed.

04

Drops out, never fights

Burner management and flame-scanner trips are never touched.

The layer drops out on any runback or trip, and comes back only when an operator says so.

  • The plant dials out

    The server inside the plant opens the connection outward. Zero inbound firewall ports into the control network.
  • Two historians

    Your live historian stays in the control network. A mirror sits in the DMZ, and every analytic query hits the mirror.
  • One write path, on site

    Biases go from an on-site server to a whitelist of control-system registers. Nothing outside the plant can write.
  • A diode where it matters

    Reads from turbine governors or safety systems leave through a one-way optical link. Nothing can travel back.

The first twelve weeks, and after.

People go to the plant.

  1. 01Weeks 1 to 3

    Walk and audit

    Walk the plant. Audit every instrument we will rely on. Align lab timestamps to process time.
  2. 02Weeks 4 to 8

    Shadow

    The model runs connected to nothing. Every prediction is scored against what actually happened.
  3. 03Weeks 9 to 12

    Advisory

    Recommendations reach the control room inside agreed limits. The operator signs each one.
  4. 04After week 12

    Closed loop, if earned

    Small trims on agreed setpoints, inside DCS clamps, revertible with one switch.

Most boiler losses are set by your coal. Three are not.

Mill bias. Worn mills get an equal fuel share and the fireball goes lopsided. Setting each mill by its condition is worth 0.1 to 0.5 point of boiler efficiency.

Soot blowing. Fixed eight-hour cycles blow clean tubes and leave dirty ones waiting. Blowing each bank only when its cleanliness factor falls means 15 to 30% fewer blows, worth 2 to 4 kcal/kWh.

On record, control moves 8 to 25 kcal/kWh. Anyone promising 50 has not run a unit. We plan on 10.

  • Unit size

    500 MW
  • Plant load factor

    70%
  • Coal, as received

    3,800 kcal/kg
  • Coal saved per year

    8,070 tonnes
  • Worth per year

    Rs 1.8 crore at a pithead station to Rs 4.9 crore far from the mine
Coal saved = heat rate gain × MW × 8,760 × PLF ÷ coal GCV. Landed coal prices from CERC orders. Range of gains: Sargent and Lundy; a US DOE-funded trial reached about 16 kcal/kWh.

Why every kiln is burned hotter than it needs to be.

Free lime, every minute. Read from torque, pyrometer, shell scanner and NOx, and corrected by every lab result at the moment its sample was taken.

The refractory, read from outside. The shell scanner you own, read for lining thickness and bound by heat physics. Plan stops on the brick that is left, not the calendar.

  • Line

    10,000 tonnes per day, 330 days
  • Clinker per year

    3.3 million tonnes
  • Petcoke saved

    4,125 tonnes a year
  • Worth per year

    Rs 5.3 to 5.8 crore
Saving = specific heat gain × clinker × fuel cost per 1,000 kcal. The best published kiln optimiser result is 20 kcal/kg; we plan on 10. Benchmarks: CII Energy Benchmarking for Cement.

A tonne in the silo is cash in the silo.

Freight now costs as much as power and fuel in a tonne of cement, and the average load travels hundreds of kilometres.

Dispatch

Two layers, because the plan breaks by 10 am.

A 24 to 72 hour plan, re-decided in seconds when a loader breaks.

2 to 5% off freight per tonne-kilometre and 10 to 25% less gate dwell.

Reconciliation

Two systems that should agree about money, checked on every row.

Toll tags against axle counts, weighbridge against invoice, coal received against coal billed.

Every disagreement gets a second look, and a person reviews every evidence file before a dispute.

Your ERP

When AI touches your ERP, it reads. A named person writes.

Every change passes format checks, business rules and a dry run, and is signed by an approver before it posts, under their name.

Wherever a plant, a line or a fleet turns cost into output.

Thermal power

Heat rate, mills, soot blowing.

Cement

Free lime, refractory, dispatch.

Steel and metals

Furnaces, staves, cranes.

Oil, gas and refining

Drilling time, units, utilities.

Manufacturing and OEMs

Lines, quality, suppliers.

Precision engineering

Machines, tolerances, scrap.

Aviation and automobile

Assembly, maintenance, parts.

Supply chain and warehousing

Inventory, dispatch, ports.

From the furnace to the port.

Manufacturing lines

Scrap, changeovers, and the line that always runs slow.

Defects are caught at the station that made them, changeovers are sequenced by what they cost, and yesterday's scrap is traced to its cause before today's shift starts.

  • Cameras find defects as they happen
  • Batch quality traced back through the process
  • Schedules that re-plan when a machine stops
Operational Intelligence
Aviation and precision engineering

Tolerances in microns. Delays in days.

Assembly, maintenance and parts on one model, so a missing part is seen before the aircraft is waiting for it, and a drifting tolerance is caught before it becomes rework.

  • Parts and work orders on the same object
  • Drift in machines caught before scrap
  • Maintenance planned on condition, not the calendar
Ports, rail and dispatch

A tonne in the silo is cash in the silo.

Dispatch planned 24 to 72 hours ahead across trucks, rakes and barges, then re-decided in seconds when a loader breaks.

  • 2 to 5% off freight per tonne-kilometre
  • 10 to 25% less gate dwell
  • Every disagreement about money gets a second look

No plant-wide digital twin. No plant-wide camera analytics.

A whole-plant simulation drifts with every batch, and plant-wide camera pilots get muted.

Questions to ask anyone who pitches you AI for your plant.

Including us.

Who on your team has run a unit, and for how long?
A good answer: names, plants and years on shift.
How will the saving be measured?
A good answer: on-and-off tests over weeks, corrected for fuel quality and load, against a baseline frozen in writing before the pilot starts.
Split your heat-rate gain into its parts.
A good answer: a loss table that adds up (oxygen, unburnt carbon, spray, air heater exit), inside the published range of 8 to 25 kcal/kWh.
What does it do at 55% load, on five mills, with the grid ramping?
A good answer: specific drop-out logic. Not “the model adapts”.
Who configures the clamps, and has the control-system supplier signed off?
A good answer: your C&I team, under management of change, with a letter from the supplier.
What happens after an overhaul: new rolls, seals, burner tips?
A good answer: a defined degraded mode, and what retuning costs.

Foundry for Industry, plainly.

What does Foundry for Industry change in a plant?
Three numbers the plant is already measured on: cost per tonne (fuel, power and freight), variance from shift to shift, and cash tied up in inventory and unreconciled bills. It does not try to sell more tonnes; volume is set by the market.
Can it trip the unit?
It is designed not to. It sits above the control system, never inside it, and never writes to a valve position, a drive speed, a trip or an interlock. Safety systems stay on their own controllers. Clamps and rate limits are configured in your control system by your own team.
How does data leave the plant?
The plant dials out. No session from IT or the cloud reaches the control network, a historian mirror in the DMZ answers every query, and reads from turbine governors or safety systems leave through a one-way data diode.
Which industries is it for?
Thermal power, cement, steel and metals, oil, gas and refining, manufacturing, OEMs and suppliers, precision engineering, aviation and automobile, and the supply chain around them: warehousing, inventory, dispatch and ports.
How long until it is running?
Twelve weeks from first walk-through to advisory recommendations in the control room. Closed-loop operation comes after that, only if it has been earned in advisory mode.

Read access to your historian. An instrument audit. Recommendations your operators sign.