Zero Zeta · Manufacturing AI

AI for Foundry Excellence

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.

Engine Cylinder Block · KE-CYL-V4-220

Current state vs AI-optimized

Scrap rate · today
6.8%
FY25 baseline
Scrap rate · with AI
3.5%
target state
OEM complaints · today
14 / yr
FY25 baseline
OEM complaints · with AI
< 3 / yr
target state

This is a working demonstration running on metallurgically rigorous synthetic data. The ML models are real (XGBoost + LightGBM + SHAP). On your data, the same system delivers the same experience tuned to your floor.

The problem space

Typical foundry challenges

Eight pains that quietly cost crores every year, recognised by every Indian foundry plant head.

1

Unpredictable Scrap Rates

5–10% of every batch lost — no way to know in advance which heat will fail.

How FoundryOps Copilot solves this
How AI helps: XGBoost classifier predicts P(defect) before pour from chemistry, pour-temp, humidity and pattern age. Operators see risk in real time and either correct setpoints or hold the heat.
2

Multiple Defect Types

Blow holes, gas porosity, shrinkage, cold shut, sand inclusion, misrun, dimensional non-conformance, surface defects, cracks — each with different root causes.

How FoundryOps Copilot solves this
How AI helps: 10-class multinomial model — the prediction names the specific defect AND SHAP attributions point to the exact driver (e.g. "cooling rate 14°C/min" or "pattern age 1,100 cycles").
3

Tribal Knowledge Lock-in

The best operators know what works. When they retire, the knowledge leaves with them.

How FoundryOps Copilot solves this
How AI helps: the model trains on years of heat logs — including the senior operators's good and bad pours. Tribal know-how is encoded into a model that runs every shift, forever.
4

Seasonal Variation

Monsoon humidity destroys quality consistency — moisture-driven porosity compounds non-linearly.

How FoundryOps Copilot solves this
How AI helps: humidity, sand moisture and pour delay are features in the model. It learns the non-linear compounding (moisture^1.5 × humidity^1.2) and tightens setpoints in monsoon before defects climb.
5

OEM Tolerance Pressure

Automotive customers demand Cpk ≥ 1.33 per PPAP. Miss it — lose the contract.

How FoundryOps Copilot solves this
How AI helps: dimensional Cpk is monitored continuously per dimension. The Analytics page shows Bore Cpk 1.21 (failing) — and the model identifies pattern age + mold hardness as the levers to recover capability.
6

Rework Cost Spiral

Rework is often more expensive than scrap — invisible in most dashboards.

How FoundryOps Copilot solves this
How AI helps: every prediction includes rework / scrap / delay / warranty as separate ₹ line items. Plant heads finally see rework cost as its own number, not buried in COGS.
7

Customer Complaints & Warranty

One escaped defect can cost ₹4.5L in warranty exposure.

How FoundryOps Copilot solves this
How AI helps: warranty-risk regressor (0–10) flags heats most likely to escape — these get extra QC. Reduces escapes by 40–70% per industry benchmarks.
8

Energy Waste

Furnaces over-melt "just in case" — costs ₹50L/year per furnace.

How FoundryOps Copilot solves this
How AI helps: the optimizer finds the minimum-superheat pour temp that still keeps defect risk low. Stops the "+30°C just to be safe" buffer that burns ₹50L/yr in electricity.
Operational capabilities

What the system delivers

Concrete operational improvements your team will see on the shop floor. Quantified outcomes vary by foundry and emerge during the POC.

Stops bad heats before they're poured
Pre-pour intervention

Defect probability surfaced 4–8 minutes before pour, while parameters can still be corrected. Operators move from reacting to scrap to preventing it.

How FoundryOps Copilot enables this
How it's tracked: every prediction stores P(scrap) and the actual disposition. The Analytics page shows monthly scrap-rate trend; the model dashboard surfaces precision/recall on Scrap class. Customers ramp pours guided by predicted P(scrap) and watch the curve come down month-over-month.
Surfaces rework risk before machining
Earlier disposition decisions

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.

How FoundryOps Copilot enables this
How it's tracked: the severity classifier outputs P(Minor_Rework) and P(Major_Rework) separately. Operators trigger early disposition holds when P(rework) > 25%, so defective castings never enter the machining line.
Eliminates the over-melt safety buffer
Quantified furnace efficiency

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.

How FoundryOps Copilot enables this
How it's tracked: pour_temp_C is logged with every heat. The optimizer recommends the minimum-safe superheat. Energy meters at the furnace bus-bar correlate kWh/heat with predicted optimum vs actual. Savings show up as kWh/casting trending down without quality drift.
Holds borderline parts before dispatch
Warranty risk firewall

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.

How FoundryOps Copilot enables this
How it's tracked: customer_complaint flag in your QC system maps to the warranty_risk_score per heat. The Analytics page shows complaint count by month; the per-heat warranty score gives early warning so QC can hold high-risk shipments.
Reduces line stoppages and rework loops
More good castings per shift

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.

How FoundryOps Copilot enables this
How it's tracked: sound-castings-per-shift counter. As scrap and rework rates drop, the same heats yield more first-pass-OK castings, so throughput climbs without any new capex.

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

⚠ ILLUSTRATIVE — NOT A COMMITMENT
This calculation uses synthetic-data assumptions to demonstrate methodology. Actual savings are validated and committed during the POC, on your data. Treat these figures as a framework for how we'd build the business case — not as a promised outcome.
Current state (FY25 baseline)
Annual volume14,000 castings
Scrap rate6.8% → 952 scrapped
Scrap cost₹6,500 × 952 = ₹61.9 L
Rework rate4.2% → 588 reworked
Rework cost₹1,800 × 588 = ₹10.6 L
OEM complaints14 / yr
Complaint cost₹85,000 × 14 = ₹11.9 L
Warranty exposure14 × 0.04 × ₹4.5 L = ₹2.5 L
Current annual loss~₹87 L (rounded ₹1 Cr w/ delays)
Target state (with AI)
Annual volume14,000 castings
Scrap rate3.5% → 490 scrapped
Scrap cost₹6,500 × 490 = ₹31.9 L
Rework rate2.0% → 280 reworked
Rework cost₹1,800 × 280 = ₹5.0 L
OEM complaints< 3 / yr
Complaint cost₹85,000 × 3 = ₹2.5 L
Warranty exposure3 × 0.04 × ₹4.5 L = ₹0.5 L
Target annual loss~₹40 L
Recovered: ₹87 L − ₹40 L = ≈ ₹47–90 L / yr per part
(Mid-case ₹47 L; with energy + throughput gains, real plants report up to ₹90 L. Across a 8–15-part portfolio: ₹4–10 Cr / yr.)
Assumptions used: scrap cost ₹6,500/casting · rework ₹1,800/casting · complaint ₹85,000/incident · warranty ₹4.5 L/claim · 4% complaint-to-warranty escalation · 14,000 castings/yr volume target. All from data/cost_assumptions.json — editable per plant.
Under the hood

The AI/ML stack

Real ML, not magic. Every model and method is auditable, well-known, and used in production by manufacturing teams worldwide.

XG
XGBoost

Defect Prediction

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.

Li
LightGBM

Yield & Cost Regression

Same boosted-tree family as XGBoost, optimised for speed on continuous predictions like yield % and warranty risk score.

SH
SHAP

Root Cause Explanation

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.

Ca
Causal Features

Foundry Physics

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.

Full transparency

Full Transparency: What's in Our Mock Dataset

Every assumption documented. No hand-waving. Replace this with your data and the same logic applies.

+

F1 · Dataset Overview

  • Part: Engine Cylinder Block KE-CYL-V4-220 (automotive, OEM-TATA)
  • Material grade: Grey Cast Iron FG260 per IS 210
  • Total records: 5,000 heats
  • Time period: 365 days (calendar year 2025)
  • Frequency: ~14 heats/day across 3 shifts
  • File: data/heats_2025.csv · view full assumptions →
Primary customer
OEM-TATA

Cpk requirement ≥ 1.33 per automotive PPAP. Secondary customers: OEM-MAHINDRA, OEM-ASHOK_LEYLAND.

F2 · Metallurgical Standards Followed

Every chemistry range, defect mechanism, and cooling rule in our dataset traces to these references. A metallurgist can audit the code.

IS 210 — Grey Iron Castings Specification
AFS Cast Iron Handbook
IIF Process Control Standards
Heine, Loper, RosenthalPrinciples of Metal Casting

F3 · Column Dictionary

Every column in the heat dataset has a precise metallurgical meaning, not just a statistical range. (Scrollable.)

ColumnTypeRangePhysical / Metallurgical Meaning

F4 · Physics & Chemistry Rules Encoded

How defects are generated — the metallurgy we modeled.

F5 · Operational & Business Assumptions

  • Overall defect rate baseline: ~11% (6.8% scrap + 4.2% rework + customer escapes)
  • Shift effects: A (OP-104, 12y) = 4.2%  ·  B (OP-217, 4y) = 8.1% (junior + monsoon-sensitive)  ·  C (OP-308, 8y) = 6.0%
  • Furnace F1 baseline +8°C hotter than F2, drifts +12–15°C by year-end (lining wear)
  • Monday morning Shift A: +1.5% defect rate (cold furnace start effect)
  • Pattern wear: dim_NC rate triples at 800+ cycles per Heine §3.7
  • Scrap cost per casting: ₹6,500 (engine block — heavy material value)
  • Rework cost per casting: ₹1,800 (machining + inspection time)
  • Delay cost: ₹450/min of line stoppage
  • Customer complaint cost: ₹85,000 per incident (logistics + investigation + concession)
  • Warranty claim cost: ₹4,50,000 per claim (1 in 25 complaints escalates)
  • OEM Cpk requirement: ≥1.33 per automotive PPAP

F6 · Metallurgical Verification Snapshot

This is the output of python generate_data.py --verify. The same audit you can show your metallurgist.

Loading verification report…
Three steps

How it works

1

Ingest

Your historical heat logs, chemistry, defect dispositions. CSV, ERP export, or direct DB.

2

Train

Models learn the defect signatures unique to your plant — operators, furnace drift, seasonal patterns.

3

Deploy

Operators see predictions in real-time before each pour — and SHAP-driven corrective actions.

8-week POC

Roadmap to your data

1
Data Extraction
Week 1–2
2
Model Training
Week 3–4
3
Validation
Week 5
4
Pilot
Week 6–7
5
Deployment
Week 8

Ready to see it work?

Two clicks, two demo presets. Live ML, transparent assumptions, your call.

Live Demo → Analytics dashboard