Custom algorithm production

XINHUOAI Custom Model Training Platform

Self-service custom AI model training platform for uploading field materials, running GPU training, evaluating models, and exporting deployable models to AI cameras, AI edge boxes, and AI NVRs.

Request a Demo View Products

XINHUOAI turns vertical AI requests into a repeatable model-production workflow. Instead of treating every new target as a one-off software project, the customer can upload field materials, create a training task, evaluate results, and export the model to local AI devices.

What It Solves

Many industrial scenes need targets that are not in generic AI libraries: special equipment states, warning signs, workwear colors, dedicated vehicles, object absence, wrong placement, or site-specific actions. The platform helps convert those needs into model assets.

Data Quality Rules

Useful training materials should cover real lighting, camera angle, distance, occlusion, normal samples, abnormal samples, and edge cases. Field data is better than generic photos because deployment quality depends on matching the actual scene.

Model Deployment

The platform is for model production. After the model is exported, real-time detection, alarm judgment, snapshots, and evidence records run locally on the AI camera, AI edge box, or AI NVR.

Specifications

Key technical information.

WorkflowTask creation, material upload, dataset cleaning, annotation, GPU training, evaluation, model export

RuntimeAfter export, models run on local AI devices without external internet for daily recognition

Billing modelTraining time can be measured by GPU usage, with optional annotation, tuning, and deployment services

Target devicesAI cameras, AI edge boxes, AI NVRs, and compatible edge analytics products

Custom model work

What is required before a custom video model is trained

Custom training is most useful when a site needs to recognize a specific object, item of PPE, vehicle condition, work action, label, or operating state that is not covered by a standard event model. The work starts with a defined visual task and representative footage from the intended camera view.

Write the visual decision in plain language

State what should be detected, what should be ignored, which camera views are in scope, and what evidence would be sufficient for an event. A request such as identify unsafe loading is too broad until it is expressed as visible actions, objects, positions, or time conditions that can be reviewed from video.

Collect positive and difficult negative examples

Useful training material includes normal examples, target examples, and scenes that look similar but should not raise an alarm. Different shifts, lighting, camera angles, uniforms, weather, equipment types, partial occlusion, and background changes should be represented where they occur in the real deployment.

Keep model version and rule version separate

The model identifies visual features; the site rule decides when that observation becomes an alarm. Recording the model version, confidence setting, region, duration, schedule, and event interval makes it possible to understand why a result changed after an update.

Validate on held-out site footage

Before wider use, test the deployed model on images or clips that were not used for training. Review missed detections, false alarms, low-confidence cases, and evidence quality with the people who understand the operation. The resulting sample set can support later improvement work.

Related planning: Guide: multimodal review of uncertain alarms. A custom model should be evaluated against the actual camera view and the operational decision it will support.

Project inquiry

Tell us your camera environment and required AI analytics.

For the first version this site uses a lightweight email-based inquiry flow. It avoids a database and keeps the English site simple until inquiry volume justifies a dedicated CRM integration.