Back to Case Studies

Rack-Inspector: AI Pallet-Rack Safety Inspection with Human-in-the-Loop Grading

Executive Summary

GenAI Protos built Rack-Inspector, a working end-to-end MVP that turns a single warehouse rack photo into defect boxes, a GREEN / AMBER / RED safety grade, and an audit-ready inspection record. It supports three regional standards (SEMA / EN 15635, ANSI MH16.1, AS 4084), catches 87% of critical RED cases, costs under $0.01 per photo, and returns results in about 3 seconds. Qualified inspectors stay in control of every decision that carries safety or legal weight.

01Client Requirement

Every pallet rack is a load-bearing steel structure, and a single forklift strike, missing locking pin, detached brace, or missing anchor can turn a safe bay into a collapse risk. Manual inspection is slow, subjective, and inconsistent: two inspectors can grade the same bay differently, and the paper trail is often thin when an incident is investigated. The requirement was an AI-assisted inspection workflow that would: Flag visible rack damage from a field photo and point inspectors directly to defects Apply the correct traffic-light grade for each region's safety standard Never auto-clear a hazard, and keep humans on millimetre-level and legally meaningful calls Produce a defensible, audit-ready record for every inspection

Section image

02Our Approach

We started with the data. Analysis of 200 real inspection photos plus a blind feasibility test (which reached 87%) confirmed the approach was viable and shaped two key decisions: millimetre boundaries stay with humans and gauges, and inspector annotations drawn on photos are excluded from training so the model learns rack condition, not markup. We then built a layered detect, grade, apply-rules, and route-to-human pipeline: Capture and triage: A fine-tuned CNN ensemble (two EfficientNetV2-S variants plus ConvNeXt-Tiny) returns a fast GREEN / AMBER / RED guess, used only as a weak prior. VLM defect grading: Gemini (via OpenRouter) localises defects, draws boxes, estimates region-neutral severity, and explains its reasoning as validated JSON. Region rule layer: The only layer that assigns colour. It maps neutral severity to each region's thresholds, action wording, and cited clause, with categorical RED rules for missing pins, anchors, or bracing. Human-in-the-loop gate: RED, AMBER, low-confidence, poor-photo, and millimetre-borderline cases are routed to a qualified inspector. Audit report and feedback loop: Every output stores defect boxes, grade, clause, and raw evidence, and every confirmed human decision feeds back as clean training data. Deployment was split across tiers: a lean FastAPI app on Railway, the classifier on an on-prem GPU service, and the VLM as a managed API. If the classifier fails or slows down, the system degrades gracefully to "no hint" rather than blocking or silently clearing a hazard.

03Outcome

Next phases cover a live warehouse pilot with guided capture and photo-quality gating (with zero missed REDs as the bar), followed by on-device grading, offline sync, and multi-site rollout.

  • 87% of critical RED cases caught, graded serious, or flagged for sign-off
  • ~92% recall on the critical RED class from the classifier ensemble
  • Under $0.01 cost per photo graded
  • ~3 seconds mean latency per photo
  • 3 regional standards supported (UK/EU, US, AU) from a single core model
  • A defensible audit trail with defect boxes, cited clauses, raw evidence, and human confirmations
  • A safety-first design that never silently clears a hazard

Frequently asked questions

Rack-Inspector is an AI-assisted pallet-rack safety inspection system built by GenAI Protos. It converts a single warehouse rack photo into defect boxes, a GREEN / AMBER / RED safety grade, and an audit-ready inspection record.

No. The AI flags and localises damage, but every RED, AMBER, low-confidence, poor-photo, and millimetre-borderline case is routed to a qualified human inspector. The system is designed for accuracy and efficiency, not autonomy.

It supports SEMA / EN 15635 (UK/EU), ANSI MH16.1 (US), and AS 4084 (Australia). A region rule layer maps the same detected defect to the correct grade, action wording, and cited clause for each standard.

In testing, the system caught 87% of critical RED cases, and the classifier ensemble achieved about 92% recall on the critical RED class. Accuracy improves further as confirmed field inspections feed into the active-learning loop.

Yes. The same pattern (AI detection, a transparent rules layer, a human review gate, an audit record, and active learning) applies to industrial inspection, facility safety, asset condition review, maintenance triage, and compliance workflows.

Build AI safety inspection workflows that stay defensible.

Tell us what you want to build, test, or improve. We'll help you define a practical first step.

Let's Explore Your Idea