Question 1

What is Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Production PatchCore anomaly detection system on WideResNet-50 — trains exclusively on normal casting images, builds a CoreSet memory bank of normal patch features, and flags deviations at inference with anomaly score heatmaps via a Django operator UI.

Question 2

What problem does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection solve?

Accepted Answer

Industrial foundries produce casting parts that must meet strict quality standards. Defects — porosity, cracks, surface roughness, inclusions — can cause part failures downstream. Manual visual inspection is slow, inconsistent (especially at end of shift), expensive, and hard to scale.

Traditional supervised defect classifiers need thousands of labeled defect images per defect type — difficult i…

Question 3

Why was Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection built?

Accepted Answer

Supervised vision models fail in foundries because defect examples are scarce by design — you cannot collect thousands of crack images without scrapping production. PatchCore's unsupervised approach treats deviation from normal as the signal itself.

I built a Django-wrapped system so QC operators could upload castings, see heatmap overlays, and confirm rejections without any ML expertise — while…

Question 4

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection work at a high level?

Accepted Answer

Built PatchCore on WideResNet-50 in a Django app: CoreSet memory bank trained on normal castings only; inference uses nearest-neighbor distances to score anomalies and render defect heatmaps for operators.

Question 5

What was the business impact of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

~0.94 AUROC on real foundry images with zero defect labels at train time. Detects novel defect types, localizes regions via heatmaps, and lets non-ML operators run inspections from a web UI.

Question 6

What type of project is Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection is a Computer Vision project delivered in 2025, with AI Engineer ownership across architecture and implementation.

Question 7

What was AI Engineer's role on Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Implemented PatchCore training and inference pipeline with WideResNet-50 backbone and CoreSet memory bank subsampling. Built Django dataset manager: image upload, OK/DEFECT labeling, train/test split, and management commands for preparation. Delivered inspection UI with anomaly score display, heatmap overlay on original image, and inspection history. Authored evaluation suite measuring AUROC, pix…

Question 8

What technologies power Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection uses PyTorch, PatchCore, WideResNet-50, Django, Python 3.11, Bootstrap, SQLite, PIL, OpenCV, torchvision, FAISS, CUDA, and additional production tooling.

Question 9

What AI capabilities are included in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Dhruvil Patel built a Foundry Defect Detection system at 1POINT1 using PatchCore on WideResNet-50 — training exclusively on normal casting images with CoreSet memory bank subsampling, achieving ~0.94 AUROC, spatial anomaly heatmaps, and a Django operator UI for non-ML QC staff.

Question 10

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection relate to Dhruvil Patel's portfolio?

Accepted Answer

Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection demonstrates production AI engineering in Computer Vision, with measurable outcomes and documented architecture.

Question 11

Can I see a live demo of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

No. Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection was built under client confidentiality. Live previews and screenshots are omitted due to client policies and NDAs.

Question 12

Where is the source code for Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Implementation detail is covered in the approach, architecture, and tradeoffs sections of this case study.

Question 13

What happens in approach step 1 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Feature extraction: feed OK casting images through pretrained WideResNet-50; extract feature maps at layers 2 and 3 for rich surface texture and edge representations.

Question 14

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Feature extraction: feed OK casting images through pretrained WideResNet-50; ex…?

Accepted Answer

Feature extraction: feed OK casting images through pretrained WideResNet-50; extract feature maps at layers 2 and 3 for rich surface texture and edge representations.

Question 15

What happens in approach step 2 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Patch aggregation: aggregate local feature patches from each image — each patch represents a small spatial region of the casting surface.

Question 16

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Patch aggregation: aggregate local feature patches from each image — each patch…?

Accepted Answer

Patch aggregation: aggregate local feature patches from each image — each patch represents a small spatial region of the casting surface.

Question 17

What happens in approach step 3 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

CoreSet subsampling: greedy selection reduces ~120,000 training patches to ~5,000 representative patches while preserving normal-appearance coverage.

Question 18

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: CoreSet subsampling: greedy selection reduces ~120,000 training patches to ~5,0…?

Accepted Answer

CoreSet subsampling: greedy selection reduces ~120,000 training patches to ~5,000 representative patches while preserving normal-appearance coverage.

Question 19

What happens in approach step 4 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Memory bank persistence: save compact normal patch memory bank to disk — the model's knowledge of defect-free casting appearance at every position.

Question 20

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Memory bank persistence: save compact normal patch memory bank to disk — the mo…?

Accepted Answer

Memory bank persistence: save compact normal patch memory bank to disk — the model's knowledge of defect-free casting appearance at every position.

Question 21

What happens in approach step 5 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Inference: extract patches from new image, FAISS nearest-neighbor distance to memory bank per patch, aggregate to image-level anomaly score.

Question 22

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Inference: extract patches from new image, FAISS nearest-neighbor distance to m…?

Accepted Answer

Inference: extract patches from new image, FAISS nearest-neighbor distance to memory bank per patch, aggregate to image-level anomaly score.

Question 23

What happens in approach step 6 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Heatmap generation: color each region by local anomaly score (red = high anomaly) for operator review and defect localization.

Question 24

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Heatmap generation: color each region by local anomaly score (red = high anomal…?

Accepted Answer

Heatmap generation: color each region by local anomaly score (red = high anomaly) for operator review and defect localization.

Question 25

What happens in approach step 7 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Threshold calibration: optimal threshold balanced on held-out labeled test set (OK + known defective) after training.

Question 26

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Threshold calibration: optimal threshold balanced on held-out labeled test set…?

Accepted Answer

Threshold calibration: optimal threshold balanced on held-out labeled test set (OK + known defective) after training.

Question 27

What happens in approach step 8 of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Clean architecture: ml/ (pure Python, no Django), inspection/ (Django views), core/ (shared utilities) — ML testable and camera-integrable independently.

Question 28

How does Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection handle: Clean architecture: ml/ (pure Python, no Django), inspection/ (Django views), c…?

Accepted Answer

Clean architecture: ml/ (pure Python, no Django), inspection/ (Django views), core/ (shared utilities) — ML testable and camera-integrable independently.

Question 29

What challenge 1 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

Training only on normal data — no defect examples available

Question 30

How was "Training only on normal data — no defect examples available" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

PatchCore unsupervised anomaly detection: deviation from the normal patch distribution is the defect signal — no labeled defects required.

Question 31

What challenge 2 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

Memory bank would contain millions of patches from training images

Question 32

How was "Memory bank would contain millions of patches from training…" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Greedy CoreSet subsampling reduces ~120,000 patches to ~5,000 representative patches while preserving coverage of normal appearances.

Question 33

What challenge 3 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

Threshold selection without labeled defects during training

Question 34

How was "Threshold selection without labeled defects during training" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Calibrate optimal threshold on a small held-out labeled test set (OK + known defective) after training completes.

Question 35

What challenge 4 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

Varying image quality from foundry inspection cameras

Question 36

How was "Varying image quality from foundry inspection cameras" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Preprocessing pipeline: resize to 256×256, normalize, and center-crop to standard input regardless of source camera.

Question 37

What challenge 5 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

Not all foundries provide CSV metadata with images

Question 38

How was "Not all foundries provide CSV metadata with images" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

CSV-independent pipeline — images organized by directory structure alone (train/good, test/good, test/defective).

Question 39

What challenge 6 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection face?

Accepted Answer

ML code tightly coupled to Django would block camera integrations

Question 40

How was "ML code tightly coupled to Django would block camera integr…" solved in Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Strict ml/ vs inspection/ separation — ML modules have zero Django imports and run in batch or camera scripts independently.

Question 41

What result 1 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection achieve?

Accepted Answer

Achieved ~0.94 AUROC on held-out test set with real foundry casting images (100 OK + 50 known defective).

Question 42

What result 2 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection achieve?

Accepted Answer

Zero defect labels required for training — model learns exclusively from normal casting appearances.

Question 43

What result 3 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection achieve?

Accepted Answer

Detects novel defect types not present in training data, including porosity, cracks, and inclusions.

Question 44

What result 4 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection achieve?

Accepted Answer

Spatial heatmap localizes defect regions within the image for operator confirmation before rejection.

Question 45

What result 5 did Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection achieve?

Accepted Answer

Clean ML/Django separation enables integration with camera feeds, video frames, and batch processing pipelines.

Question 46

What did the team learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (insight 1)?

Accepted Answer

Unsupervised anomaly detection is the right fit when defects are rare by design — supervised models cannot learn what they never see enough of.

Question 47

What did the team learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (insight 2)?

Accepted Answer

CoreSet subsampling is essential for PatchCore at production scale; the full patch bank is not practical for FAISS search.

Question 48

What did the team learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (insight 3)?

Accepted Answer

Heatmaps are as important as the score — operators need to see where the anomaly is, not just a pass/fail number.

Question 49

What did the team learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (insight 4)?

Accepted Answer

Separating ml/ from Django early pays off when camera integrations and batch scripts need the same inference core.

Question 50

What did the team learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (insight 5)?

Accepted Answer

Threshold calibration on a small labeled eval set is acceptable even when training is fully unsupervised.

Question 51

Key takeaway 1 from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

PatchCore trains on OK images only — zero defect labels required

Question 52

Key takeaway 2 from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

~0.94 AUROC on real foundry casting test images

Question 53

Key takeaway 3 from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Detects novel defect types via deviation from normal patch distribution

Question 54

Key takeaway 4 from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Spatial heatmaps localize defects for operator confirmation

Question 55

Key takeaway 5 from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

ml/ module separated from Django for camera and batch integration

Question 56

How did AI Engineer contribute to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (item 1)?

Accepted Answer

Implemented PatchCore training and inference pipeline with WideResNet-50 backbone and CoreSet memory bank subsampling.

Question 57

How did AI Engineer contribute to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (item 2)?

Accepted Answer

Built Django dataset manager: image upload, OK/DEFECT labeling, train/test split, and management commands for preparation.

Question 58

How did AI Engineer contribute to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (item 3)?

Accepted Answer

Delivered inspection UI with anomaly score display, heatmap overlay on original image, and inspection history.

Question 59

How did AI Engineer contribute to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (item 4)?

Accepted Answer

Authored evaluation suite measuring AUROC, pixel-level AUROC, and optimal threshold calibration.

Question 60

How did AI Engineer contribute to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection (item 5)?

Accepted Answer

Separated ML core from Django orchestration for batch processing and live camera feed integration paths.

Question 61

How was Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection built?

Accepted Answer

Built PatchCore on WideResNet-50 in a Django app: CoreSet memory bank trained on normal castings only; inference uses nearest-neighbor distances to score anomalies and render defect heatmaps for operators.

Question 62

What was the outcome of Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

~0.94 AUROC on real foundry images with zero defect labels at train time. Detects novel defect types, localizes regions via heatmaps, and lets non-ML operators run inspections from a web UI.

Question 63

What did AI Engineer learn from Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Key learnings include Unsupervised anomaly detection is the right fit when defects are rare by design — supervised models cannot learn what they never see enough of., CoreSet subsampling is essential for PatchCore at production scale; the full patch bank is not practical for FAISS search., Heatmaps are as important as the score — operators need to see where the anomaly is, not just a pass/fail number., Separating ml/ from Django early pays off when camera integrations and batch scripts need the same inference core., Threshold calibration on a small labeled eval set is acceptable even when training is fully unsupervised..

Question 64

What was AI Engineer's contribution to Foundry Defect Detection: Unsupervised Anomaly Detection for Industrial Casting Inspection?

Accepted Answer

Dhruvil Patel contributed across Implemented PatchCore training and inference pipeline with WideResNet-50 backbone and CoreSet memory bank subsampling., Built Django dataset manager: image upload, OK/DEFECT labeling, train/test split, and management commands for preparation., Delivered inspection UI with anomaly score display, heatmap overlay on original image, and inspection history., Authored evaluation suite measuring AUROC, pixel-level AUROC, and optimal threshold calibration., Separated ML core from Django orchestration for batch processing and live camera feed integration paths., Built preprocessing pipeline: resize to 256×256, normalize, center-crop for varying foundry camera quality., with ownership from design through deployment.