canonical: https://jentic.com/apis/aws.amazon.com/aws-lookout-for-vision # AWS Amazon Lookout for Vision API Amazon Lookout for Vision is the AWS service for detecting visual defects in industrial products using computer vision. The API lets developers create projects, manage training and test datasets, train and host anomaly-detection models, run on-demand inference, and package models for edge deployment. The spec exposes 22 operations across 13 paths, secured by AWS Signature Version 4 in the Authorization header, with a regional base URL pattern of https://lookoutvision.{region}.amazonaws.com. ## For AI agents Train and run computer-vision anomaly-detection models on industrial imagery, then deploy them to the cloud or to edge devices for defect detection. ## Scope Does not handle general-purpose image classification, OCR, face recognition, or video analytics - use for industrial anomaly and defect detection on still images only. ## Capabilities - Create and manage Lookout for Vision projects that scope training and inference resources - Build and update training and test datasets from labelled images of normal and anomalous parts - Train, list, and describe anomaly-detection models per project version - Start and stop a model to make it available for synchronous inference - Run defect detection on a new image via the DetectAnomalies endpoint - Package a trained model for edge deployment on AWS IoT Greengrass devices ## Use cases ### Production-Line Defect Detection Manufacturers deploy Lookout for Vision to flag defective parts on the production line by sending captured images to DetectAnomalies and acting on the response. Operators train a model on labelled normal and anomalous images, host it as a project version, and integrate the inference endpoint into the line's PLC or edge device, replacing manual visual inspection that struggles with throughput and consistency. Example prompt: Call POST /2020-11-20/projects/{projectName}/models/{modelVersion}/detect with the image bytes for a part captured on the line and return the IsAnomalous flag and confidence ### Model Lifecycle Management ML engineers manage the project, dataset, and model lifecycle through the Lookout for Vision API: create projects, attach training and test datasets, kick off training jobs, and start or stop hosted models to balance inference availability against cost. The API surfaces every step of CreateProject, CreateDataset, CreateModel, StartModel, and StopModel programmatically. Example prompt: Create a project named 'pcb-quality', attach training and test datasets from S3, start training, and poll DescribeModel until status reaches HOSTED ### Edge Deployment Packaging When inference must run on-device with low latency, teams package a trained Lookout for Vision model through the model-packaging-jobs endpoint for AWS IoT Greengrass deployment. The packaged model runs locally on factory hardware while training and updates remain centralised in AWS. Example prompt: Call the model-packaging endpoint for project {projectName} model version {modelVersion} with a Greengrass component spec, then poll the packaging job status until SUCCEEDED ### Agent-Driven Quality Inspection via Jentic AI agents embedded in operations dashboards call Lookout for Vision through Jentic to surface inspection results, trigger model retraining when drift is detected, and orchestrate dataset updates as new examples are labelled. Jentic exposes the 22 Lookout for Vision operations as discoverable tools so the agent picks DetectAnomalies, StartModel, or CreateModel based on the intent. Example prompt: Given an alert 'inspection accuracy dropped on line 3', search Jentic for 'create dataset Lookout for Vision', upload a new labelled batch, and trigger retraining of the affected project's model ## Key endpoints | Method | Path | Description | | --- | --- | --- | | GET | /2020-11-20/projects | List Lookout for Vision projects | | POST | /2020-11-20/projects/{projectName}/datasets | Create a dataset under a project | | POST | /2020-11-20/projects/{projectName}/models | Train a new model version | | GET | /2020-11-20/projects/{projectName}/models/{modelVersion} | Describe a model version | | POST | /2020-11-20/projects/{projectName}/models/{modelVersion}/detect | Run anomaly detection on an image | | GET | /2020-11-20/projects/{projectName}/modelpackagingjobs/{jobName} | Describe a model packaging job | ## Key resources - **Projects** — Top-level container for datasets and models - **Datasets** — Training and test datasets of labelled images - **Models** — Trained anomaly-detection model versions and their hosting state - **DetectAnomalies** — Synchronous inference endpoint for running predictions on a new image - **Model Packaging Jobs** — Package models for AWS IoT Greengrass edge deployment ## Why Jentic - **Setup:** Wiring Amazon Lookout for Vision by hand means computing an AWS Signature Version 4 on every request from your access key and secret, selecting the correct regional host, and threading project and model-version path segments into each call. Through Jentic you install once, import the Lookout for Vision API from the API Directory, store the AWS access key and secret once, and your agent calls it. - **Permission scoping:** Lookout for Vision carries the project name and model version as URL path parameters, such as /2020-11-20/projects/{projectName}/models/{modelVersion}/detect, so a rule can pin the agent to one project and model. You choose which operations it may call, such as running defect detection, so training a model or deleting a dataset is not included unless you add it. - **Credential handling:** Your AWS access key, secret key, and optional session token are stored once, encrypted, by your own Jentic One instance and injected at execution time as the SigV4 signature. They never enter the agent's prompt, logs, or context. - **Discovery method:** Agents search Jentic by intent such as 'detect anomalies on an image' or 'check a model version', and Jentic returns the matching Lookout for Vision operation with its input schema, including the project and model-version path parameters, so the agent calls the right endpoint without hand-building the request. ## Related APIs - **Google Cloud Vision** — Google Cloud Vision offers general-purpose image classification and object detection without Lookout for Vision's industrial-anomaly focus. - **Google Cloud AutoML Vision** — AutoML Vision lets teams train custom image classifiers with similar low-code training flow but without integrated edge packaging for industrial use. - **Cloudinary Upload API** — Use Cloudinary to manage and serve the source images that feed into Lookout for Vision training and inference. ## FAQ ### What authentication does the Amazon Lookout for Vision API use? Lookout for Vision uses AWS Signature Version 4. The Authorization header on each request must be a SigV4 signature derived from the IAM access key, secret key, and optional session token. Through Jentic, AWS credentials live in your Jentic One instance and Jentic constructs the SigV4 signature for the agent on each call. ### Can I run defect detection on an image with the Lookout for Vision API? Yes. POST /2020-11-20/projects/{projectName}/models/{modelVersion}/detect runs synchronous inference on the image bytes you supply and returns whether the part is anomalous, the confidence score, and (for segmentation models) anomaly masks per defect type. The model must be in HOSTED state before calling DetectAnomalies. ### What are the rate limits for the Amazon Lookout for Vision API? Lookout for Vision applies per-account, per-region request quotas published in the AWS service quotas console. DetectAnomalies typically supports tens of transactions per second per hosted model, with higher limits available on request. Throttling returns HTTP 400 with a ThrottlingException - back off using exponential retry. ### How do I start a hosted model through Jentic? Search Jentic for 'start a Lookout for Vision model', load the schema for POST /2020-11-20/projects/{projectName}/models/{modelVersion}/start, and execute. After the call, poll GET /2020-11-20/projects/{projectName}/models/{modelVersion} until status is HOSTED before invoking DetectAnomalies. ### How is Lookout for Vision priced? AWS bills Lookout for Vision based on training hours, hosted-inference hours per active model, and DetectAnomalies image volume. Stopping a hosted model with the StopModel operation halts the inference-hour charges - agents that only run inference during shifts should stop the model overnight to save cost. ### Can I deploy Lookout for Vision models to edge devices? Yes. The model-packaging-jobs endpoints let you package a trained model as an AWS IoT Greengrass component so inference runs on-device for low-latency factory environments. Training and dataset management remain in AWS while inference happens locally. ### Can I limit what my agent is allowed to do with the Amazon Lookout for Vision API? Yes. Because Jentic One is self-hosted, your own rules decide which Lookout for Vision operations and AWS credentials the agent may use. Since the project name and model version travel as URL path parameters, such as /2020-11-20/projects/{projectName}/models/{modelVersion}/detect, you can pin the agent to a single project and model. You also choose which operations it may call, so you can allow running DetectAnomalies while excluding training a model or deleting a dataset unless you add them.