Every pour carries risk. Your data can reduce it.
Zero Zeta FoundryOps Copilot uses your historical process and quality data to predict
casting defects before metal is poured — enabling smarter decisions that improve
yield, reduce rejection, and protect customer quality. Engineered for automotive and engine
castings, calibrated to IS 210 FG260 grey iron.
Eight pains that quietly cost crores every year, recognised by every Indian foundry plant head.
5–10% of every batch lost — no way to know in advance which heat will fail.
Blow holes, gas porosity, shrinkage, cold shut, sand inclusion, misrun, dimensional non-conformance, surface defects, cracks — each with different root causes.
The best operators know what works. When they retire, the knowledge leaves with them.
Monsoon humidity destroys quality consistency — moisture-driven porosity compounds non-linearly.
Automotive customers demand Cpk ≥ 1.33 per PPAP. Miss it — lose the contract.
Rework is often more expensive than scrap — invisible in most dashboards.
One escaped defect can cost ₹4.5L in warranty exposure.
Furnaces over-melt "just in case" — costs ₹50L/year per furnace.
Concrete operational improvements your team will see on the shop floor. Quantified outcomes vary by foundry and emerge during the POC.
Defect probability surfaced 4–8 minutes before pour, while parameters can still be corrected. Operators move from reacting to scrap to preventing it.
Severity classifier separates likely scrap, likely rework, and likely good — at the heat level. Stops marginal castings from entering machining loops where rework cost compounds.
Energy regressor models kWh per ton from process inputs. Reveals how much "just in case" superheat your team adds and recommends tightened windows backed by defect-risk constraints — safe to apply.
Composite warranty risk score (0–10) flags any heat scoring above 7 for additional inspection. Borderline parts get caught at the gate, not at the OEM's incoming inspection — or worse, in the field.
Fewer scrapped pours means fewer mid-shift resets. Earlier disposition decisions mean machining centers run cleaner. Anomaly detection surfaces equipment drift before it becomes downtime.
Quantified business impact for your portfolio is developed during the 8-week POC, calibrated to your actual production volumes, unit costs, and historical defect patterns. ▾ See illustrative calculation
| Annual volume | 14,000 castings |
| Scrap rate | 6.8% → 952 scrapped |
| Scrap cost | ₹6,500 × 952 = ₹61.9 L |
| Rework rate | 4.2% → 588 reworked |
| Rework cost | ₹1,800 × 588 = ₹10.6 L |
| OEM complaints | 14 / yr |
| Complaint cost | ₹85,000 × 14 = ₹11.9 L |
| Warranty exposure | 14 × 0.04 × ₹4.5 L = ₹2.5 L |
| Current annual loss | ~₹87 L (rounded ₹1 Cr w/ delays) |
| Annual volume | 14,000 castings |
| Scrap rate | 3.5% → 490 scrapped |
| Scrap cost | ₹6,500 × 490 = ₹31.9 L |
| Rework rate | 2.0% → 280 reworked |
| Rework cost | ₹1,800 × 280 = ₹5.0 L |
| OEM complaints | < 3 / yr |
| Complaint cost | ₹85,000 × 3 = ₹2.5 L |
| Warranty exposure | 3 × 0.04 × ₹4.5 L = ₹0.5 L |
| Target annual loss | ~₹40 L |
data/cost_assumptions.json — editable per plant.
Real ML, not magic. Every model and method is auditable, well-known, and used in production by manufacturing teams worldwide.
Gradient-boosted decision tree classifier. Imagine 500 expert metallurgists each asking different yes/no questions about your heat parameters, then voting. Industry standard for tabular ML. Probability outputs in 50 ms.
Same boosted-tree family as XGBoost, optimised for speed on continuous predictions like yield % and warranty risk score.
SHapley Additive exPlanations — from cooperative game theory. For every prediction, tells you which parameters pushed the model toward "defect" and by how much. No black box.
We encode foundry physics: carbon equivalent, thermal gradients, pattern wear curves, monsoon adjustment factors. The model learns from data; the features come from metallurgical science.
Every assumption documented. No hand-waving. Replace this with your data and the same logic applies.
data/heats_2025.csv · view full assumptions →Cpk requirement ≥ 1.33 per automotive PPAP. Secondary customers: OEM-MAHINDRA, OEM-ASHOK_LEYLAND.
Every chemistry range, defect mechanism, and cooling rule in our dataset traces to these references. A metallurgist can audit the code.
Every column in the heat dataset has a precise metallurgical meaning, not just a statistical range. (Scrollable.)
| Column | Type | Range | Physical / Metallurgical Meaning |
|---|
How defects are generated — the metallurgy we modeled.
This is the output of python generate_data.py --verify. The same audit you can show your metallurgist.
Loading verification report…
Your historical heat logs, chemistry, defect dispositions. CSV, ERP export, or direct DB.
Models learn the defect signatures unique to your plant — operators, furnace drift, seasonal patterns.
Operators see predictions in real-time before each pour — and SHAP-driven corrective actions.
Two clicks, two demo presets. Live ML, transparent assumptions, your call.