Rack-Inspector: AI Pallet-Rack Safety Inspection

Rack-Inspector addresses a high-stakes warehouse safety problem: every pallet rack is a load-bearing steel structure, and a forklift strike can quietly turn a safe bay into a collapse risk. The industry inspection model uses a GREEN / AMBER / RED traffic-light grade under standards such as SEMA / EN 15635, ANSI MH16.1, and AS 4084, where the colour carries operational and legal meaning.

The system keeps human judgement where it belongs while making the inspection workflow faster, more consistent, and more auditable. A single field photo flows through a detect, grade, apply-rules, and route-to-human pipeline. AI flags visible damage and localises defects; the rule layer maps severity to the relevant regional standard; and a qualified inspector confirms cases that carry risk, require millimetre measurement, or have low confidence.

The current build is a working end-to-end MVP using FastAPI, Gemini VLM, a CNN ensemble, and a React interface. The approach was shaped by a real inspection dataset of 200 photos and a blind feasibility test, with evidence that the main delivery work is better capture, clean data, precise grading rules, and a defensible human-in-the-loop workflow rather than unsupported autonomy.

Domain

Warehouse pallet-rack safety.

Standards supported

SEMA / EN 15635, ANSI MH16.1, and AS 4084.

Core stack

FastAPI, Gemini VLM, CNN ensemble, and React.

Status

Working end-to-end MVP.

Inspection output

Defect boxes, GREEN / AMBER / RED grade, region action wording, cited standard clause, raw evidence, and audit-ready record.

Design principle

Accuracy and efficiency over autonomy. AI assists; qualified humans remain on decisions that carry safety and legal weight.

Safety routing

RED, AMBER, low-confidence, millimetre-borderline, and poor-photo cases are flagged for human review.

Measured performance

87% of critical RED cases caught, under $0.01 cost per photo, about 3 seconds mean latency, and 3 regional standards supported.

Every pallet rack in a warehouse is a load-bearing steel structure. Damage from a forklift strike, missing locking pin, detached bracing, missing anchor, corrosion, or bend can turn a normal bay into a serious safety risk. The established inspection process depends on trained people walking aisles and assigning traffic-light grades.

Done by hand, the process is slow, subjective, and inconsistent. Two inspectors can grade the same bay differently, and the paper trail can be thin when an incident is investigated. The opportunity is to keep qualified human judgement in the process, while using AI to flag visible damage, point the inspector directly to potential problems, and preserve a stronger audit record.

Before building the grader, the inspection dataset was analysed across 200 real photos plus a blind feasibility test. The analysis confirmed that the approach is feasible and shaped two core design decisions: millimetre-level boundaries must stay with humans and gauges, and training should focus on the steel condition rather than inspector annotations drawn onto images.

The Rack-Inspector system was built as an AI-assisted inspection workflow that turns a single field photo into a structured safety record. It detects and localises visible defects, estimates severity in region-neutral terms, applies a standards-specific rule layer, and produces colour-coded outputs for qualified human review where neede

The implementation combines a fast capture and triage classifier, a Gemini vision-language model grader for substantive defect reasoning, a region rule layer for GREEN / AMBER / RED assignment, a human-in-the-loop gate for risky or uncertain cases, and an audit-defensible report that keeps raw evidence and reasoning attached to the decision.

The design deliberately separates AI severity estimation from regional colour assignment. The model detects defects and estimates severity in neutral terms, while a data-driven configuration maps those outputs to each region's thresholds, colour scheme, action wording, and cited clause. That lets the same detected defect produce the correct region-specific verdict across UK/EU, US, and AU deployments.

The architecture separates image capture, AI grading, region-specific rule application, human review, audit reporting, and feedback. This keeps the safety decision traceable: the AI proposes, the rule layer applies the standard, and the human gate handles cases where risk, low confidence, poor image quality, or millimetre-level judgement is involved.

Capture and triage classifier

A field photo enters the workflow, and a fine-tuned CNN ensemble returns a fast GREEN / AMBER / RED guess as a weak prior

VLM grader

Gemini via OpenRouter identifies visible defects, draws boxes, estimates region-neutral severity, and explains reasoning as strict validated JSON.

Region rule layer

The only layer that decides colour. It maps neutral severity to the deployment standard and adds categorical RED rules for missing pin, anchor, or brace cases.

Human-in-the-loop gate

Flags RED, AMBER, low-confidence, millimetre-borderline, and poor-photo cases for qualified review.

Audit-defensible report

Returns colour-coded defect boxes, rack RAG grade, region action wording, cited standard clause, and raw evidence as JSON and UI output.

Feedback loop

Every confirmed human decision becomes clean training data for active learning.

The inspection workflow follows the source system design: one field photo enters the pipeline, automated components flag visible damage and propose structured evidence, region rules decide the colour, and qualified human review is triggered wherever the decision carries safety or legal weight.

This section captures the system decisions that make the MVP useful for a real safety inspection process: region-neutral AI grading, rule-based colour assignment, safe human gating, graceful degradation, and audit-ready reporting.

Capture and triage

A fine-tuned CNN ensemble provides a fast GREEN / AMBER / RED hint that is treated only as a weak prior.

VLM defect reasoning

Gemini VLM localises defects, draws boxes, rates severity in neutral terms, and explains the reasoning in validated JSON.

Rule-based grading

A region rule layer maps neutral severity to standards-specific colours, actions, thresholds, and cited clauses.

Human review design

RED, AMBER, low-confidence, poor-photo, and millimetre-borderline cases are routed to qualified inspection.

Graceful degradation

The classifier is best-effort. If unavailable or slow, the inspection continues with the VLM and no classifier hint

Audit report

The system returns colour-coded boxes, rack grade, action wording, cited standard clause, raw evidence, and JSON output

Active learning

Confirmed human labels feed back into the dataset to improve future model performanc

The technical details below preserve the source evidence from the rough case study, including dataset size, feasibility results, model performance, deployment footprint, latency, cost, standards coverage, and human-review routing rules.

Rack inspection is a safety and legal process, not a pure computer vision exercise. The value of the system is that it improves speed, consistency, and auditability while preserving qualified human judgement for decisions that carry legal weight.

Rack-Inspector is a working end-to-end MVP that demonstrates a defensible design for AI-assisted pallet-rack safety inspection. It makes inspection faster and more consistent, catches obvious damage automatically, routes uncertainty to humans, and produces a richer evidence record than a paper-based process.

This case study can be reused as a pattern for safety inspection systems where AI can accelerate evidence capture and triage, but humans must remain responsible for legally meaningful decisions. The same approach can apply to industrial inspection, facility safety, asset condition review, maintenance triage, and compliance workflows.

• AI Detection Layer: Use computer vision to locate visible defects and estimate severity in neutral terms.
• Rules Layer: Keep formal decisions, thresholds, colour grades, and clauses in a transparent configuration layer.
• Human Review Gate: Route risky, uncertain, borderline, or low-quality inputs to a qualified reviewer.
• Audit Record: Preserve evidence, boxes, scores, rationale, action wording, and human confirmations.
• Active Learning Loop: Turn confirmed field decisions into better training data over time.