Incoming Quality Prediction
“Will this incoming batch pass inspection?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Will this incoming batch pass inspection?
100% incoming inspection is expensive and slow. Skip-lot strategies save time but miss quality escapes that cause downstream defects costing 10-50x the incoming material value. For an automotive manufacturer receiving 5,000 batches per month, predicting which batches need full inspection vs. skip-lot saves $12M in inspection costs while reducing quality escapes by 70%, preventing $35M in warranty claims.
Quick answer
Graph neural networks predict incoming batch quality before inspection by learning patterns across supplier history, material specifications, production run data, and inspection records. The model detects supplier process drift and spec trends that skip-lot sampling misses, saving $47M annually for an automotive manufacturer: $12M in inspection costs and $35M in prevented warranty claims.
Approaches compared
4 ways to solve this problem
1. Skip-lot sampling (AQL-based inspection)
Inspect a random sample from each batch based on Acceptable Quality Level tables. Reduce inspection frequency for suppliers with good track records.
Best for
Reduces inspection costs when supplier quality is consistently good. Industry standard with clear statistical basis.
Watch out for
Random sampling misses systematic defects. If a supplier's process drifted last week, the defective batch passes through with the same probability as any other. Quality escapes cost 10-50x the incoming material value in downstream rework and warranty claims.
2. Statistical process control (SPC) on supplier metrics
Track supplier defect rates and material spec measurements on control charts. Increase inspection frequency when metrics go out of control.
Best for
Detects systematic shifts in supplier quality over time. Well-understood in manufacturing quality management.
Watch out for
Reactive -- you detect the shift after defective batches have already arrived. Control charts also monitor single metrics independently, missing the compound pattern of multiple specs drifting simultaneously.
3. Classification model on batch features (XGBoost)
Train a classifier on features like 'supplier defect rate trend,' 'time since last audit,' and 'spec measurements from certificates of analysis' to predict pass/fail per batch.
Best for
Predictive rather than reactive. Can flag high-risk batches before they arrive. Captures the most important single-table signals.
Watch out for
Treats each batch independently. Cannot model how a supplier's quality correlates with their production equipment age, their sub-tier material sources, or how their quality interacts with specific material-spec combinations.
4. KumoRFM (relational graph ML)
Connect suppliers, materials, inspections, production runs, and specs into a quality graph. The GNN learns which supplier-material-production combinations produce defects and detects process drift before it causes quality escapes.
Best for
Highest accuracy for directing inspection resources. Detects compound quality risks: supplier process change + material spec drift + equipment stress. Reduces quality escapes by 70% while cutting inspection costs.
Watch out for
Requires supplier, material, and inspection data in normalized tables. Most impactful when you have 50+ suppliers and detailed inspection records with spec measurements.
Key metric: Automotive manufacturers save $47M annually ($12M inspection + $35M warranty prevention) by predicting incoming batch quality and directing inspection to high-risk batches.
Why relational data changes the answer
Quality is a multi-table problem. Whether a batch passes inspection depends on the supplier's process stability (tracked in quality records over time), the material specification's sensitivity to process variation (stored in specs), how the supplier's equipment has been performing (inferred from production runs), and whether the raw materials from their sub-tier supplier have shifted (reflected in certificate-of-analysis measurements). Skip-lot sampling ignores all of this context. SPC monitors metrics independently. Even a good classification model on batch features misses the cross-table interactions.
Relational models connect the full quality graph. They learn that MAT202 from PlastiForm Co has a 72% fail probability because their defect rate has trended up 280% over three months, their quality tier dropped from A to B, the Melt Flow Index in their last three batches is trending toward the upper spec limit, and the production runs using this material have been yielding 95.2% quality instead of the typical 99%+. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table approaches. For quality prediction, that gap means catching 70% of quality escapes that would otherwise become $35M in warranty claims.
Skip-lot sampling is like a food safety inspector who checks every 10th restaurant delivery at random. They catch obvious contamination but miss the supplier whose refrigeration truck has been running 3 degrees warm for a week, whose last two batches had elevated bacteria counts just below the threshold, and who just switched to a cheaper raw ingredient source. Graph-based quality prediction reads the full supplier health record and directs the inspector to check the at-risk deliveries.
How KumoRFM solves this
Graph-powered intelligence for supply chains
Kumo connects suppliers, materials, inspections, production runs, and specs into a quality graph. The GNN learns which supplier-material-production combinations produce defects: when a supplier's process changed, when raw material specs drifted, or when a production run stressed equipment beyond normal parameters. PQL predicts pass/fail per incoming batch, directing inspection resources to high-risk shipments.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
SUPPLIERS
| supplier_id | name | quality_tier | last_audit |
|---|---|---|---|
| SUP101 | MetalWorks Inc | A | 2025-01-15 |
| SUP102 | PlastiForm Co | B | 2024-11-20 |
| SUP103 | ChemSynth Ltd | A+ | 2025-02-10 |
MATERIALS
| material_id | name | spec_range | supplier_id |
|---|---|---|---|
| MAT201 | Steel Alloy 7075 | Tensile: 570-590 MPa | SUP101 |
| MAT202 | ABS Resin Batch | MFI: 18-22 g/10min | SUP102 |
| MAT203 | Epoxy Compound | Viscosity: 4000-5000 cP | SUP103 |
INSPECTIONS
| inspection_id | material_id | result | defect_rate | date |
|---|---|---|---|---|
| INS301 | MAT201 | Pass | 0.2% | 2025-02-15 |
| INS302 | MAT202 | Fail | 4.8% | 2025-02-20 |
| INS303 | MAT203 | Pass | 0.1% | 2025-02-22 |
PRODUCTION_RUNS
| run_id | material_id | equipment_id | output_quality |
|---|---|---|---|
| RUN401 | MAT201 | EQ01 | 99.8% |
| RUN402 | MAT202 | EQ02 | 95.2% |
| RUN403 | MAT203 | EQ03 | 99.9% |
SPECS
| material_id | parameter | target | tolerance |
|---|---|---|---|
| MAT201 | Tensile Strength | 580 MPa | +/- 10 MPa |
| MAT202 | Melt Flow Index | 20 g/10min | +/- 2 |
| MAT203 | Viscosity | 4500 cP | +/- 500 cP |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(INSPECTIONS.result = 'Fail', 0, 1, days) FOR EACH MATERIALS.material_id
Prediction output
Every entity gets a score, updated continuously
| MATERIAL_ID | SUPPLIER | FAIL_PROB | RECOMMENDED_ACTION |
|---|---|---|---|
| MAT201 | MetalWorks Inc | 0.05 | Skip-lot OK |
| MAT202 | PlastiForm Co | 0.72 | Full inspection |
| MAT203 | ChemSynth Ltd | 0.02 | Skip-lot OK |
Understand why
Every prediction includes feature attributions — no black boxes
MAT202 -- ABS Resin Batch from PlastiForm Co
Predicted: 72% fail probability (Full inspection recommended)
Top contributing features
Recent inspection failure rate trend
+280% (3 months)
30% attribution
Supplier quality tier deterioration
A to B
24% attribution
MFI spec drift in last 3 batches
Trending high
20% attribution
Production run quality with this material
95.2%
15% attribution
Time since last supplier audit
4 months
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about incoming quality prediction
How do you predict incoming batch quality before inspection?
Connect supplier quality history, material specifications, inspection records, and production data in a graph model. The model learns which supplier-material combinations are trending toward defects and flags high-risk batches for full inspection before they arrive. This is predictive rather than reactive -- you catch quality issues before they escape into production.
How do you reduce quality escapes in manufacturing?
Direct inspection resources based on predicted risk rather than random sampling. Graph models identify the 15-20% of incoming batches that carry 80% of the defect risk, enabling full inspection where it matters and skip-lot where quality is stable. This reduces quality escapes by 70% while actually lowering total inspection costs.
What data do you need for quality prediction?
Supplier profiles with quality tier and audit history, material specifications with tolerance ranges, incoming inspection records with defect rates and measurements, and production run quality data. For best results, add certificate-of-analysis data from suppliers and sub-tier supplier information. Each additional data source improves early detection of process drift.
What is the cost of a quality escape in manufacturing?
Quality escapes cost 10-50x the incoming material value in downstream rework, scrap, production delays, and warranty claims. A $5 defective part that makes it into an assembled product can generate $250+ in warranty costs. For an automotive manufacturer, preventing quality escapes saves $35M in annual warranty claims alone.
How do you detect supplier process drift early?
Monitor the compound pattern across multiple quality dimensions simultaneously. A single metric drifting (e.g., spec trending toward the upper limit) may not be alarming alone. But when spec drift combines with increasing defect rates, longer time since last audit, and production quality declining, the convergence signals process drift that will produce defective batches within weeks.
Bottom line: An automotive manufacturer saves $47M annually ($12M in inspection costs + $35M in prevented warranty claims) by predicting incoming batch quality. Kumo's quality graph detects supplier process drift and material spec trends that skip-lot sampling misses.
Related use cases
Explore more supply chain use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
