Use Case · Manufacturing & Industrial AI

AI-assisted quality inspection for manual manufacturing lines.

DehazeLabs helps manufacturers turn manual visual inspection into defect detection, grading support, operator review queues, and production quality analytics. The workflow runs from image capture through labeled dataset to operator dashboard — starting narrow, proving value, then expanding. Lighthouse pilot in leather manufacturing; the same pipeline applies across textiles, electronics, automotive materials, packaging, and logistics QC.

→ The problem with manual inspection at scale

Manual visual inspection has three failure modes that compound as production volume grows. First, inconsistency: experienced inspectors disagree on borderline defects — and that disagreement is invisible until a customer complaint surfaces it. Second, no analytics: defects are tallied on paper or spreadsheets, so the pattern data that would identify upstream process problems never reaches the people who can fix them. Third, throughput pressure: when production accelerates, inspection time per unit compresses, and escape rate increases.

AI-assisted inspection doesn't replace the inspector. It gives them a prioritized review queue — images ranked by predicted defect probability and severity — so their attention goes to the units that most need human judgment. It also creates the structured defect dataset that makes batch-level analytics and continuous improvement possible for the first time.

The 4-week lighthouse workflow

W1
Discovery + data collection
Workflow map · defect taxonomy · 500–2,000 sample images · capture protocol · inter-inspector agreement baseline · batch metadata schema
W2
Labeling + dataset setup
CVAT annotation campaign · FiftyOne dataset load · data quality report (lighting, blur, class balance) · capture protocol refinement
W3
Baseline model + evaluation
Defect classifier or detection model · FiftyOne evaluation view · failure-case slices · false positive/negative analysis · model improvement roadmap
W4
Operator review dashboard
Image · predicted defect · grade outcome · confidence · severity · batch · review status · approve/reject flow · batch quality report

Defect taxonomy — start narrow

The taxonomy is defined in Week 1 after seeing real samples. Starting with 5–8 classes is correct. Expanding before you have sufficient labeled examples per class is the common mistake that degrades model performance. A simplified example for surface-material inspection:

scratch
Linear surface mark, depth and length determine severity
discoloration
Uneven pigmentation, chemical staining, fading
wrinkle
Permanent crease or fold; distinguished from handling marks
stain
Localized contamination; oil, dye bleed, handling residue
cut / tear
Structural break; auto-reject in most grading systems
edge damage
Fraying, cracking, or deformation at material edges
grain inconsistency
Defined precisely before labeling — subjective class, high labeler disagreement risk
acceptable
No defect or within tolerance; required for classifier calibration

Taxonomy is refined after first data review. Do not finalize before seeing real samples from the line.

What the operator dashboard shows

The dashboard is not a model output viewer. It is a review queue — the interface through which the inspector's attention is directed to the units that most need human judgment, and through which their decisions are captured and made queryable.

→ Operator review queue — example row
image_idbatch_042 · unit_0187 · station_2
predicted_defectscratch · localization: center-left
confidence0.87
severitymedium — grade impact likely
grade_recommendationrework / inspect under raking light
review_statuspending review

Grade outcome, not just defect type

The dashboard captures both the predicted defect label and the inspection decision: acceptable, rework, or reject. This is what makes batch-level quality analytics possible — not just "how many scratches" but "which batches produced the most reworks, and what do those units have in common." That pattern data is what connects the inspection workflow to upstream process improvement.

Capture protocol — the thing that determines everything

Bad image capture ruins model performance more reliably than any model choice. The Week 1 facility visit produces a standardized capture protocol: lighting type and angle (raking light for surface defects; diffuse overhead washes out scratches), camera angle and working distance, background material, sample handling to prevent adding handling marks, and image naming convention that encodes batch and station metadata.

For most manual inspection stations, a standardized capture setup costs under $500 in materials. Investing in the protocol in Week 1 determines whether the labeled dataset is usable in Week 2.

→ Lighthouse case: Maryadha
We built the first lighthouse pilot with Maryadha — a leather manufacturing network operating factories across India. Manual inspection of hides for scratches, wrinkles, discoloration, stains, cuts, and edge damage. Scope: a single inspection station, 500+ sample images, defect taxonomy defined with the inspection team, a labeled FiftyOne dataset, a baseline classifier, and an operator review dashboard in 4 weeks. Maryadha's network gives us real factory access and a real production problem. The results — and the workflow — generalize directly to any surface-material inspection: textiles, electronics, automotive, packaging, logistics QC. This is how Manufacturing Quality AI becomes a reusable practice, not a leather-specific solution.
[ What changes ]

From paper tally sheets to a queryable defect dataset.

Inspection consistency
Measured and improvable
Inter-inspector agreement captured in Week 1. Ambiguous defect classes identified before labeling. Grading standards codified in the defect taxonomy.
Inspector attention
Directed by model confidence
Review queue surfaces the units most likely to need human judgment. High-confidence clean units move through faster. Borderline cases get the scrutiny they warrant.
Batch analytics
Available for the first time
Defect type, severity, and grade outcome structured and queryable by batch, station, and time period. The data that connects inspection to upstream process decisions.
Case study
Reusable proof
Anonymized pilot results demonstrate the workflow to manufacturing buyers across industries. One lighthouse pilot becomes the proof that scales the practice.
→ Pilot timeline
Week 0 Facility selected · champion named · data/case-study permission agreed · pilot price agreed · discovery visit scheduled Agreement
Week 1 Facility visit · workflow map · defect taxonomy · 500–2,000 sample images · capture protocol · inter-inspector agreement baseline Current-state map · sample dataset · taxonomy v1
Week 2 CVAT annotation · FiftyOne dataset load · data quality report · capture protocol refinement · class balance analysis Labeled dataset v1 · FiftyOne view · data quality report
Week 3 Baseline model train/fine-tune · FiftyOne evaluation · failure-case slices · false positive/negative analysis Baseline results · evaluation dashboard · improvement roadmap
Week 4 Operator review dashboard · approve/reject flow · batch quality report · pilot readout document Dashboard · batch report · production proposal
[ FAQ ]

Manufacturing Quality AI — common questions.

Does this replace human inspectors?
No. AI-assisted inspection helps quality teams review faster, find patterns, and standardize defect classification — it does not replace inspector judgment. The operator review queue is designed to surface the cases most worth a human's attention, not to automate away the inspection function. Human sign-off on reject decisions remains in the workflow.
What materials and product types does this work for?
The workflow applies to any product with visually distinguishable defect types: leather, textiles and apparel, electronics PCBs and components, automotive materials and trim, food and beverage packaging, ceramics, glass, and processed metals. The defect taxonomy and image capture protocol differ by material — but the pipeline from image capture through FiftyOne curation to operator dashboard is the same.
What does a lighthouse pilot actually deliver in 4 weeks?
A 4-week lighthouse pilot delivers: a capture protocol specific to your inspection station, a labeled dataset of 500+ images covering your primary defect types, a FiftyOne dataset view for inspection and class analysis, a baseline defect detection/classification model, and an operator review dashboard showing predicted defect, confidence, severity, and batch context. It also surfaces data quality issues — inconsistent lighting, class imbalance, labeler disagreement — before you invest in a production system.
How consistent does our current inspection process need to be before this works?
It doesn't need to be consistent — but we need to measure how consistent it is. Inter-inspector agreement is part of the Week 1 discovery. If two experienced inspectors disagree significantly on which defects are acceptable, that ambiguity will surface in the labeled data and in model performance. Surfacing it is valuable: it tells you which defect classes need clearer grading standards before a model can learn them reliably.
What cameras and lighting do you need?
The lighthouse pilot starts with what's available: a phone camera, DSLR, or existing camera feed. The Week 1 discovery produces a recommended capture protocol — lighting type (raking light is standard for surface defects), camera angle, working distance, background, and sample handling. For most manual inspection stations, a basic standardized capture setup costs under $500. Production-grade camera installation is out of scope for the lighthouse pilot.

Manual inspection with no defect analytics, inconsistent grading, and growing volume?

The lighthouse pilot is designed to be fast and low-risk: narrow scope, light price, real results in 4 weeks, and an anonymized case study that justifies a production rollout. Tell us about your facility and inspection workflow — we'll tell you whether this maps to your situation.

→ Related reading

Physical AI Data Flywheel → Production CV deployment → AI for manufacturing → Edge & Perception AI → Agentic Operations →