EVA

EVA is an innovative solution that integrates Machine Learning (ML), Large Language Models (LLM),
and Vision-Language Models (VLM) to transform standard cameras into smart AI cameras.
Without complex coding, users can configure AI through natural conversations and quickly deploy AI services tailored to their environments.

Service Scenarios

Synergy of Multiple Foundation Models

EVA seamlessly integrates ML, VLM, and LLM to upgrade ordinary cameras into AI-powered cameras.
Once a camera is connected to EVA, the LLM and VLM automatically suggest the optimal models and logic.
When the ML model detects objects, the VLM analyzes the context to accurately identify risks, such as missing protective gear, and delivers alerts.
Every step can be configured through natural language conversations, with multiple agents supporting operations.

Transform into Smart AI Cameras

AI Cameras Made Easy for Everyone

Simply by entering camera network information, EVA instantly converts ordinary cameras into smart AI cameras.
It enables immediate application of scenarios such as risk detection, activity monitoring, and quality inspection.

With EVA Agent

EVA Agent: A Powerful AI Camera Companion

EVA Agent maximizes AI camera performance with minimal input.
Based on user-provided data, it executes optimal operations tailored to each situation.

AI Controlled by Conversation

EVA allows users to intuitively control AI through natural dialogue.
With just conversations, users can configure AI for camera streams and contexts with ease.

Simple On-site Customization

EVA enables vision models optimized for the field without complex coding, data collection, or labeling.
By intuitively designating targets, EVA automatically customizes the models.

Detecting Diverse Scenarios

EVA combines vision models with multimodal LLMs to perform precise video analysis.
It addresses a variety of needs such as anomaly detection, quality control, and safety monitoring,
while delivering real-time notifications to platforms like Teams and Slack via API.

AI Evolving Through Feedback

EVA incorporates user feedback to reduce false detections and continuously improve accuracy.
It learns from false detection data to make better judgments in similar situations.
This minimizes false alarms and delivers AI cameras optimized for real-world environments.

EVA Agent

EVA Agent Inspired by the Human Brain

EVA Agent is an intelligent AI service inspired by the structure of the human brain.

(1) Reasoning Model simulates the logical thinking of the frontal lobe to recommend optimal detection scenarios or models during camera registration.
(2) LLM reflects the language comprehension of the temporal lobe, enabling AI configuration through conversation.
(3) VLM mimics the visual processing ability of the occipital lobe to analyze images.
(4) Multi-modal Vector Store functions like the hippocampus, storing data to enhance RAG-based accuracy.
By integrating all these components, EVA Agent delivers optimal performance.

Service Architecture

Efficient and Robust System Design

EVA operates AI cameras efficiently by seamlessly connecting the EVA App and EVA Agent.
Depending on the user’s infrastructure, ML, LLM, and VLM are flexibly composed to apply the best-suited models.

Collaboration Between Super Agent and Sub Agent

EVA Agent processes complex requests through a Super Agent and Sub Agent structure.
The Super Agent interprets user intent, while the Supervisor selects the appropriate Sub Agent to achieve the goal.
Each Sub Agent specializes in tasks using LLM and VLM, delivering results back to the Super Agent.
The modular design of Super and Sub Agents ensures flexibility and scalability, enabling efficient handling of complex tasks.

Optimized Multi-Foundation Models

EVA flexibly configures ML, LLM, and VLM by balancing performance and cost.
The most suitable foundation model is selected depending on the detection targets, scenarios, and context of each camera.
Sub Agents employ high-performance models for one-time tasks requiring precision, such as detection enhancement or feedback generation.
Conversely, continuous monitoring tasks utilize cost-efficient models.

Efficient Resource Utilization

EVA optimizes computing resources by separating Vision, LLM, and VLM models.
Running all models in a single system consumes excessive resources and slows processing, limiting the number of supported cameras.
The separated architecture of EVA App and EVA Agent can reduce costs by up to 70% when managing around 50 cameras.

Sustainable High-Quality AI Service

EVA is designed on the integrated architecture of the Mellerikat platform.
The EVA App and EVA Agent are deployed in the service environment (cloud or on-premise).
On the Mellerikat platform, MLOps and LLMOps manage the AI model lifecycle, while Edge Conductor and Logic Deployer deliver optimal solutions.
Through this, EVA continuously receives the latest AI logic and models to ensure cutting-edge performance.