EVA

EVA (Edge Vision Agent) is an innovative tool that combines vision models with a multimodal LLM to transform ordinary cameras into smart AI cameras.
It allows users to control AI and implement field-tailored vision solutions effortlessly through natural conversations, without the need for complex coding.

Service Scenario

Perfect Harmony of Vision Models and Multimodal LLM

Edge Vision Agent delivers powerful capabilities by integrating vision models with a multimodal LLM.
The vision model analyzes images from the camera in real-time, while the multimodal LLM enables task configuration through natural user conversations.
For example, tasks such as hazard detection, activity monitoring, and quality inspection can be easily set up using simple text commands.

Deploy a vision model suited to your camera to start real-time image analysis, and use the LLM to recognize situations and define events.
Leverage the diverse technologies provided by Mellerikat to quickly and efficiently address on-site challenges.

Transform into a Smart AI Camera

Anyone Can Easily Create a Smart AI Camera

With Edge Vision Agent, you can convert an ordinary camera into a smart AI camera by simply entering its network information.
Connect various vision agents—such as hazard detection, activity sensing, or quality inspection—to Edge Vision Agent for instant implementation.

AI Tailored Through Conversation

Edge Vision Agent is designed to intuitively control AI through natural conversations.
Users can set desired tasks via dialogue and apply customized operations for each camera, creating AI tailored to specific purposes.

Domain Understanding

AI Solutions Optimized for the Field

Every field requires different domain knowledge. Edge Vision Agent enables the implementation of vision models tailored to specific sites through conversation alone, eliminating the need for complex coding, data collection, or labeling processes.
Its key advantage is accessibility, even for those without specialized expertise.

Example of Normal vs. Defective

Video on Customizing a Specific Defect Detection Model

Context-Aware

Smart AI That Understands Situations

By combining vision models with a multimodal LLM, Edge Vision Agent analyzes images and precisely interprets situational context.
This supports intelligent, field-optimized decision-making, enabling rapid resolution of issues like anomaly detection, quality inspection, and safety monitoring.

Provides notifications about the recognized situation through various channels such as Teams, Slack, email, etc.

Service Architecture

Efficient and Powerful System Structure

EVA is an AI service composed of EVA Application and EVA Multi-modal Service, providing multi-modal services by optimally utilizing resources.

EVA Application : Analyzes camera footage in real-time using Vision API, and through conversation with the LLM of EVA Multi-modal Service,
sets models suitable for the field or requests situation recognition.

EVA Multi-modal Service : Provides APIs for EVA Application to utilize multi-modal LLMs, and allows customization for desired purposes through various logics.
By isolating the resource-intensive multi-modal LLM environment, it is shared with EVA Applications as needed,
while EVA Application continuously performs image analysis in a minimal spec environment.

Based on the EVA platform utilizing various products and services from Mellerikat, EVA Application operates connected to the EVA platform.
Multiple EVA Applications are installed and operate where they can connect to cameras,
and the EVA platform supports EVA Applications to continuously provide high-quality services.
To utilize all features of EVA, installation in the cloud is recommended, and EVA Application for on-premise is planned to be supported.

In the overall architecture, the EVA platform provides AI services to EVA Application through MLOps and LLMOps systems,
continuously manages AI performance, and deploys the latest verified solutions.

EVA Controller provides functions to optimize models and infrastructure according to the desired service.
Not only customizing models for the field but also flexibly usable in various areas such as computing infrastructure management,
hardware control, inference interval setting, scheduling, data management, etc. (To be released)