FoundryOps Copilot — AI in Casting

ZeroZeta's FoundryOps Copilot An AI pour-station co-pilot for Indian foundries — predicts the defect 4 to 8 minutes before tap, names the parameter driving it, and turns every heat into a rupee line the plant head can defend.

Today, every defect is found at
shakeout — long after
the only minute that mattered.

Tomorrow, the pour-station screen becomes a second pair of eyes. Defect risk, root cause, and corrective action — minutes before tap, in the operator's vocabulary at the ladle and the plant head's vocabulary at the month-end review.

The ZeroZeta
position

Indian foundries don't need another MES dashboard or a ₹50L SAP quality module that prints month-end PDFs nobody reads. They need a co-pilot that speaks the language of the pour — defect risk, root cause, corrective action — in the four minutes that still matter, and the language of the boardroom — scrap rupees, OEM Cpk, warranty exposure — when the plant head reviews the month. One model. One screen.

15.86^{M MT}

India's annual casting output — 3rd-largest globally, ~90% MSMEs, $22B sector.

Institute of Indian Foundrymen · IFC 2026

8.5→3.5^%

Casting rejection rates achievable when predictive models replace trial-and-error tuning.

Indian Foundry Congress · 59th Congress Paper

Where FoundryOps Copilot plugs into your workflow

01 · Plan

Charge & setpoints

02 · Pour

Risk before tap

03 · Dispose

Scrap, rework, ship

04 · Ask

AI Copilot, anywhere

The FoundryOps Copilot stack — built for the way Indian foundries actually pour.

What you'll see runs on synthetic data calibrated to IS 210 FG260 grey iron — the models, SHAP explanations, and rupee math are production-grade ML. Point the same architecture at your heats and the same screen reads your plant — your operators at the ladle, your engineers at shakeout, your metallurgist on the morning round, your plant head at month-end.

Current state · Typical heat76% high risk · ₹20,248 expected loss

FoundryOps Copilot — typical heat state showing 76% defect risk, Gas Porosity at 40.5% confidence, Rework disposition, total expected loss ₹20,248 per casting, with SHAP attribution naming pour delay, humidity, and sand moisture as the top risk drivers.

The heat as your floor pours it today. Gas Porosity at 40.5% confidence, disposition Rework, yield 88.8%. SHAP names the drivers in plain language: pour delay 6.5 min, humidity 78%, sand moisture above 4.2%.

Rendered in 316 ms

AI Recommended setpoint2% low risk · ₹1,935 expected loss

FoundryOps Copilot on the AI Recommended setpoint — 2% defect risk, most-likely defect Cracks at 1.3% confidence (negligible), disposition OK, yield 98.3%, total expected loss ₹1,935 per casting.

The same heat, with corrected pour temp, delay, and chemistry. Disposition OK, yield 98.3%, severity none. Same screen, same operator — ₹18,313 of loss prevented per casting, four to eight minutes before tap.

Rendered in 94 ms

Plant analytics · last 12 monthsPareto, monsoon, Cpk in one view

FoundryOps Copilot Analytics page — Defect Pareto by cost, monthly defect rate with humidity overlay showing the monsoon porosity spike, and dimensional capability Cpk histograms for Bore Diameter, Deck Height, and Wall Thickness against the OEM PPAP threshold of 1.33.

Behind the prediction: a Pareto by cost (not count), a monsoon humidity overlay on monthly defect rate, and Cpk per critical dimension — Bore 1.21 failing, Deck 1.45 and Wall 1.62 passing the OEM-TATA PPAP threshold of 1.33.

Where every shift, furnace, & pattern lands

Built on / Plays well with your stack

XGBoost & LightGBM modelsSHAP on every predictionCalibrated to IS 210 FG260Runs on-premise

Why us — the honest version

Most foundry software is built for the audit binder and the month-end review. Ours is built for the four minutes before the metal moves.

01

Predictive, not retrospective

MES dashboards tell you what scrapped yesterday. We tell you what's about to scrap this heat — an XGBoost classifier across 10 defect classes in under 100 ms, while pour temp, delay, and sand moisture are still adjustable.

02

Every prediction is explained

Conservative foundry buyers don't trust black boxes — rightly. Every prediction ships with its top six SHAP drivers in the operator's vocabulary: "pour delay 6.5 min: +1.77", "humidity 78%: +1.25". The argument and the answer on the same page.

03

Rupees, not just percentages

A "76% defect risk" persuades no one in a CFO review. ₹20,248 of expected loss per casting — split into scrap, rework, delay, OEM complaint, and warranty reserve, each formula traceable to a calibrated cost table — is a P&L line the plant head can defend.

04

Calibrated to your plant, not a textbook

Off-the-shelf models trained on academic datasets assume a steady-state plant that doesn't exist in India. FoundryOps Copilot retrains on your heat logs — your operators, your furnaces, your patterns, your monsoon. Seasonal effects, pattern wear past 800 cycles, furnace drift — learned, not assumed.

AI Copilot — ask anything across your heat history

FoundryOps Copilot · Plant view

You Why did F1 scrap rate spike in August? Show me the heats and the driver.

Copilot F1 ran 11.4% scrap in August vs the 6.8% baseline — 47 of 412 heats. Three drivers: furnace drift hit +13 °C by Aug 15; humidity averaged 81% with sand moisture above 4.2% on most days; Pattern PT-C crossed 1,050 cycles mid-month. 31 of 47 scraps were Shift B. Want the heat table, SHAP breakdown, or reline schedule?

Or try a preset

Today's high-risk heatsCpk by part this weekWarranty-risk > 7, last 30 daysFurnace drift trend

The plant's memory, queryable.

Heat logs, chemistry, dispositions, OEM complaints, and warranty events — encoded into a model that answers in the operator's, metallurgist's, and plant head's vocabulary. Ask in English; get charts, SHAP attributions, and the underlying heats back. It works two ways:

Investigate "Why did scrap spike?", "Which shift drove the Bore Cpk drop?" — the Copilot retrieves the heats, runs SHAP, and answers in three lines.

Decide "Safest setpoint for PT-C at 1,100 cycles in monsoon?", "Hold Heat #4471 or release?" — parameters with confidence bounds, rupee impact, audit trail.

Bring us one furnace.
We'll show you the losses before they happen.

A working session on your real heat history. We ingest, retrain, and run the prediction next to what actually happened — heat by heat, rupee by rupee. Your data, your model, your weights.

Talk to us

info@zerozeta.com

Based in

Bengaluru · India