What challenges did TrueLook face before AI implementation?

Manual safety checks on large construction sites are inconsistent, hard to scale, and can miss PPE violations or unsafe-zone exposure. Traditional approaches make it difficult to maintain reliable audit trails and to monitor many sites and workers continuously.

What was the impact of the AI solution?

With the same 1,000 labeled images, the pipeline achieved mAP scores in the 80-90% range, an improvement of 20 points over an alternate provider workflow. The approach reduced training time for subsequent updates and provided governed, scalable production deployment.

Source

ExploringEvidence: Medium50/100

TrueLook AI-powered construction safety monitoring on Amazon SageMaker AI

Use case typeComputer vision monitoringUpdated Jan 9, 2026

TrueLook built an AI-powered construction safety monitoring system on Amazon SageMaker AI that automatically detects PPE and unsafe conditions from jobsite camera images. The workflow uses SageMaker Processing, SageMaker Training, SageMaker Model Registry, SageMaker Pipelines, MLflow, TensorBoard, Amazon S3, and real-time endpoints to support a multi-stage fine-tuning and active-learning loop.

Organization: TrueLook
Industry: Real Estate
Location: United States
Published: January 2026

Reported outcomes

MAP improvement: More than 20 points

Other quantified impact

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

MAP improvement: 20 points increase

AWS Machine Learning BlogJan 9, 2026Blog postExplicit claimMedium evidence strength

an improvement of 20 points over the alternate provider’s workflow

Normalized claim

MAP with 1,000 labeled images: 80-90%

AWS Machine Learning BlogJan 9, 2026Blog postExplicit claimMedium evidence strength

the pipeline achieved mAP scores in the 80–90% range

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: TrueLook
Provider: AWS
Maturity: Exploring
Linked source: AWS Machine Learning Blog

The team automated preprocessing, training, model registration, evaluation, and deployment with SageMaker Pipelines and Model Registry, then served low-latency real-time inference through managed endpoints

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

1Computer vision monitoring
2Safety monitoring

TrueLook domain-adapted a pretrained object detection model, then fine-tuned it on construction safety datasets and TrueLook-labeled images.
The team automated preprocessing, training, model registration, evaluation, and deployment with SageMaker Pipelines and Model Registry, then served low-latency real-time inference through managed endpoints.
An active-learning loop retrains as new images arrive, enabling ongoing model improvement.

Technologies

Amazon SageMaker AI Amazon S3 SageMaker Pipelines SageMaker Model Registry SageMaker Processing SageMaker Training MLflow TensorBoard YOLOv11

With the same 1,000 labeled images, the pipeline achieved mAP scores in the 80-90% range, an improvement of 20 points over an alternate provider workflow.
The approach reduced training time for subsequent updates and provided governed, scalable production deployment.

Architecture

A three-stage computer vision workflow on Amazon SageMaker AI preprocesses jobsite imagery with SageMaker Processing, trains a YOLOv11 object detection model with SageMaker Training, and version-governs approved models in SageMaker Model Registry. SageMaker Pipelines orchestrates automated evaluation, conditional promotion, and repeatable CI/CD retraining from Amazon S3 image drops, while managed real-time endpoints serve low-latency PPE detection on live video or snapshots and trigger alerts; MLflow and TensorBoard are used for experiment tracking and validation.

Sources & evidence1

Evidence: Medium50/100Evidence strength