Complete Case Study Back to Portfolio
Project 2 of 4 — Visual AI

VisionCheck AI

One photo. Instant safety report. Zero manual inspection.

The Problem & The Solution

The Problem

Workplace PPE violations go undetected between manual inspections. Human safety inspectors visit sites periodically, leaving critical violations — missing hard hats, absent fall protection, no eye gear — unaddressed for days or weeks.

OSHA fines reach $15,625 per violation. Workers get hurt.

The Solution

VisionCheck AI analyzes any workplace photo and generates a full PPE compliance report in under 15 seconds — automatically detecting violations, rating risk level, and recommending corrective actions.

No human inspector needed. No waiting.

How It Works

The system processes images in real time using a serverless orchestration pattern built on AWS, completing the entire analysis end-to-end in under 15 seconds.

Step 01 User Browser

User visits **visioncheck.buildsmartai.app** and uploads a workplace photo.

Step 02 Netlify Frontend

HTML page converts image to base64 and sends HTTP POST request.

Step 03 AWS API Gateway

Receives request, applies CORS policy, triggers Lambda function.

Step 04 AWS Lambda

Validates file type, size, and format. Orchestrates the AI call (Python 3.12).

Step 05 Amazon Bedrock

Claude Sonnet 4.5 vision model analyzes image for PPE compliance, risk level, and recommendations.

Step 06 Safety Report

Structured report displayed on screen in under 15 seconds.

AWS Tech Stack

AWS Lambda

Serverless Python function. Validates input, orchestrates AI call, handles errors. Zero idle cost.

AWS API Gateway

HTTP API endpoint. Receives image uploads, applies rate limiting and CORS. $1 per million calls.

Amazon Bedrock

Claude Sonnet 4.5 vision model. Analyzes workplace images for PPE violations in real time.

Amazon S3

Encrypted temporary image storage. Lifecycle rule automatically deletes all images after 24 hours.

AWS IAM

Least privilege role. Lambda has exactly 3 permissions — Bedrock, CloudWatch, and S3. Nothing more.

CloudWatch

Error-level logging only. Captures failures without accumulating unnecessary log costs.

What the AI Detects

PPE Items Checked

  • Hard hat / helmet
  • High-visibility safety vest
  • Protective gloves
  • Safety boots / footwear
  • Eye protection / goggles
  • Fall protection harness
🟢 LOW
🟡 MEDIUM
🔴 HIGH
⛔ CRITICAL

Production-Grade Security

🔒 IAM Least Privilege — Lambda role has exactly 3 permissions only
🔐 S3 Encryption — All images encrypted at rest using SSE-S3
Auto-deletion — Images automatically deleted after 24 hours (lifecycle rule)
🛡 Input Validation — File type, size, and format validated before any AI call
🌐 CORS Policy — API locked to authorized domain only
💰 Budget Alert — AWS billing alert set at $15/month

Key Architectural Decision

Why Bedrock over Rekognition?

Originally planned to use Amazon Rekognition's PPE Detection API. During research, discovered AWS announced Rekognition features entering maintenance mode in 2026 — directing new deployments toward Amazon Bedrock with multimodal foundation models.

Pivoted immediately to Bedrock with Claude Sonnet 4.5. This is not just a technical decision — it demonstrates understanding of where AWS is heading and building for the future, not the past.

Demo Video

See VisionCheck AI in Action

Watch VisionCheck AI analyze a real construction site photo — PPE compliance detection, violation identification, risk rating, and recommendations delivered in under 15 seconds.

Live Product

Try It Yourself

Upload any workplace or construction photo and get an instant AI safety inspection report.

Launch VisionCheck AI

Free to use · No login required · Results in under 30 seconds

Cost & Business Impact

Cost Breakdown

Lambda: Free (1M requests/month permanently)
API Gateway: ~$0.00 at portfolio usage levels
Amazon Bedrock: ~$0.03–0.08 per image analysis
S3 Storage: ~$0.01/month
Total Monthly Cost: $1–5
Zero Idle Cost — pays nothing when not in use

Business Impact

Fines Mitigations: OSHA fines reach $15,625 per violation
Inspection Frequencies: Manual audits happen weekly at best
24/7 Coverage: VisionCheck AI works continuously from any uploaded photo
Consistent Auditing: Standardized classification logic — no human bias
Predictable Scaling: Scales to analyze 1 or 1,000 images at the same cost per image
Audit Trail: Full audit logs with timestamped report exports