What's possible with AI

Four Layers of AI in FoundryOps Copilot

The defect predictor is one capability. The intelligence around it is the system. Here's how four distinct AI layers combine into a foundry operations co-pilot you can actually trust.

01

AI That Converses

"Talks back."

What it does

Every operator, plant head, and quality engineer in your foundry has questions about their data. Today, those questions take hours โ€” open SAP, export to Excel, build a pivot table, talk to the analyst, wait. FoundryOps Copilot answers them in seconds, in plain English, with the numbers cited.

Your team doesn't learn SQL. They don't learn dashboards. They just ask.

How it works

Stack: Anthropic Claude API + structured data context

User question โ†’ backend assembles a relevant data summary from the heat database โ†’ sends to Claude with a system prompt explaining the dataset โ†’ returns a natural-language answer with cited figures.

  • Response time: ~2 seconds
  • Cost: under โ‚น3 per query at scale
  • Privacy: queries can route to an on-premise Claude deployment

Example interaction

Why did Shift B have higher defects last monsoon? Which pattern is closest to retirement? Compare F1 vs F2 furnace performance What's our biggest scrap cost driver?
Compare F1 vs F2 furnace performance.
F1 is significantly underperforming F2 this year.
  • F1 defect rate: 13.2% (vs F2 at 7.6%)
  • F1 temperature drift: +14.9ยฐC by year-end (refractory wear)
  • F1 is 6 months older than F2 at year start
Recommend scheduling F1 lining replacement in the next 4โ€“6 weeks.
Reduces analyst hours per week by 60โ€“80%
Decision latency: hours โ†’ seconds
Onboards new operators 3ร— faster โ€” no dashboard training
02

AI That Watches

"Always on."

What it does

Most foundry incidents aren't catastrophic โ€” they're slow drifts. A pattern wearing past its threshold. A furnace lining thinning. An operator's pour delay creeping up by 30 seconds a week. Humans don't notice. The AI does.

An unsupervised model continuously watches every heat for unusual combinations of parameters โ€” combinations that historically correlate with defects. When something looks off, it flags the heat before pour, not after rejection.

How it works

Stack: Isolation Forest (scikit-learn)

Trained on 5,000 historical heats. Each new heat scored on a 0โ€“1 anomaly scale. Anomalies above 0.7 trigger an alert. The explanation is generated by identifying the 2 parameters that contributed most to the anomaly score.

  • Runs in 50 ms per heat
  • Zero operator workload โ€” pure background intelligence
  • No labels required โ€” unsupervised

Example output

โš  3 anomalies detected this week
HT-2025-04231Score 0.87
Pour Temp 1340ยฐC (3.2ฯƒ low)
Carbon Eq 4.45 (2.8ฯƒ high)
โ†’ Predicted defect risk: 73%
HT-2025-04102Score 0.78
Sand Moisture 5.4% (3.5ฯƒ high)
Pour Delay 9.2 min (2.9ฯƒ high)
โ†’ Predicted defect risk: 68%
HT-2025-03987Score 0.71
Cooling Rate 15.8ยฐC/min (2.6ฯƒ high)
Cr Content 0.22% (2.4ฯƒ high)
โ†’ Predicted defect risk: 54%
Catches 80%+ of drift conditions before defect formation
Zero operator effort โ€” runs continuously in background
Surfaces patterns no human dashboard can monitor
03

AI That Acts

"Decides for you."

What it does

Prediction is useful. Action is valuable. The defect predictor tells you a heat is risky. The optimizer tells you exactly what to change โ€” and proves the new setpoints will work, before you pour.

Your senior metallurgist's tribal knowledge is "if porosity is up, drop moisture and increase pour temp by 10ยฐ." The AI finds the precise, multi-parameter optimum โ€” across 12 dimensions simultaneously โ€” in 3 seconds. No trial heats. No guessing.

How it works

Stack: scikit-optimize (Bayesian optimization with Gaussian Process surrogate)

  • Objective: minimise predicted defect probability + 0.3 ร— warranty risk
  • Search space: 12 operator-controllable parameters within physical bounds
  • Iterations: 30โ€“50 simulated parameter combinations
  • Convergence: typically 3โ€“4 seconds
  • Constraints: respects physical limits (cannot ask for pour temp below liquidus)

Example optimization run

๐ŸŽฏ Optimization complete ยท 32 iterations ยท 3.1 s
Before
62%
defect risk
โ†’
After
7%
defect risk
Pour temperature1395 โ†’ 1418ยฐC
Pour delay6.5 โ†’ 1.4 min
Sand moisture5.2 โ†’ 3.3%
Mold hardness79 โ†’ 85 B
Cooling rate14.0 โ†’ 10.2ยฐC/min
Inoculant dose0.18 โ†’ 0.22%
Replaces 3 days of trial heats with a 3-second AI search
Tests 30+ combinations per call โ€” beyond human capacity
Recommendations are physics-constrained โ€” never impossible setpoints
04

AI That Knows Its Limits

"Honest about uncertainty."

What it does

The worst AI systems are confident when they shouldn't be. The best ones tell you when they don't know.

Every prediction in FoundryOps Copilot ships with a confidence interval. When the model has seen plenty of similar heats and the prediction is stable, the band is tight. When the heat is unusual โ€” extreme parameters, rare combinations โ€” the band widens and the model says so.

This is what separates an AI tool from AI snake oil.

How it works

Stack: XGBoost with prediction-interval estimation

For each prediction, we compute the entropy of the per-class probability vector. We combine that with the training-data density in the local neighbourhood of the prediction (KDE-based).

  • Tight band (ยฑ4pp) โ†’ HIGH confidence โ€” operators act on it
  • Wide band (ยฑ20pp+) โ†’ LOW confidence โ€” flag for human review

Wide intervals are not a model failure. They are the model's honesty.

Three example predictions

HEAT A HIGH confidence
Risk: 14% [12% โ€“ 16%] ยท band width 4pp
โœ“ Safe to pour
HEAT B MEDIUM confidence
Risk: 38% [27% โ€“ 49%] ยท band width 22pp
โš  Manual review recommended
HEAT C LOW confidence
Risk: 51% [22% โ€“ 80%] ยท band width 58pp
โš  Model uncertain โ€” consult metallurgist
Operators learn when to trust AI vs. when to escalate
Wide bands trigger human-in-the-loop โ€” never blind automation
Industry-standard practice โ€” used in medical AI, autonomous driving, finance

Why four layers, not one

A complete AI system isn't one capability. It's four working together.

"Conversation without action is just a chatbot.
"Monitoring without prediction is just a dashboard.
"Action without confidence is just a gamble.
"Confidence without conversation is just a black box.

FoundryOps Copilot is built around this principle: every layer reinforces every other layer. The Q&A surfaces insights the anomaly detector found. The optimizer respects the confidence intervals. The system tells you what it knows, what it suspects, and what it doesn't know.

That's the difference between AI you demo once and AI your team uses every shift.

What comes next

Beyond the four layers โ€” capability roadmap

The four layers above are live in this demo. Here's what we build next together.

Phase 2 Weeks 6โ€“12

POC Extensions

  • 7-Day Defect Forecast
    Combines weather + production schedule + pattern wear
  • Vision-Based Surface Inspection
    Camera on shakeout line, 200 ms per casting
  • Multi-Part Family Expansion
    From cylinder block to crankcase, head, housing โ€” same architecture
Phase 3 Months 3โ€“6

Production Scale

  • Predictive Maintenance
    LSTM on furnace/mold-line vibration and current
  • Energy Optimization Engine
    Charge mix + holding time optimizer
  • OEM Integration
    Automated PPAP documentation generation
Phase 4 Months 6โ€“12

Plant-Wide Intelligence

  • Multi-Plant Deployment
  • Supply Chain AI
    Charge mix optimization across vendors
  • Closed-Loop Process Control
    Direct setpoint actuation, not just recommendation

Each layer earns its place by delivering measurable value

Ready to see them work on your data?

Try the Live Demo โ†’ Scope the 8-Week POC โ†’