What pattern does the data hold?
Numerical models produce forecasts, anomaly findings and measurable outcomes.
ARG INTELLIGENCE
Intelligence,in the flow of real work.
From learning a signal on a controller to analysing equipment, retrieving enterprise knowledge and powering multimodal assistants, ARG connects intelligence to data, software and the physical world.
MACHINE LEARNINGLLM / RAGAGENTSVISION / VOICEEDGE AI
THE ARG INTELLIGENCE ARCHITECTURE01—04
ARGINTELLIGENCE
DATATime · quality · identity
KNOWLEDGEKnowledge in context
ACTIONTools & results
FEEDBACKEvaluate & learn
THE COMPLETE PATH
Observe preciselyAnalyse in contextAct with traceability
Meet the services & runtimesCONNECTED INTELLIGENCE
Four layers. One understandable path.
Select a layer to understand its role in turning data into outcomes. This is a logical map; each product’s actual connections are defined in its project architecture.
01
Observe the real world
Sensor readings, equipment history, engineering documents, enterprise records and human requests start different intelligence paths. Time, units, quality and data ownership matter as much as the values.
Data attributable to equipment, a project or a user account
What does our knowledge say?
Retrieved evidence and language models ground answers in relevant documents and data.
Which tool performs the next step?
Request context, permitted tools and service results form an explicit workflow.
INTELLIGENCE IN PRACTICE
A real question. A clear path.
Choose an application to see its inputs, processing and user outcome. These examples explain system workflows.
DIGITAL TWIN
Why is this asset consuming more than its usual pattern?
Move from a dashboard value to anomaly detection and engineering explanation, with distinct numerical-analysis and project-knowledge paths.
Explore the product- 01Input & context
Select timestamped asset history with point identity, units and data quality.
- 02Processing & analysis
ARG AI analyses consumption patterns; the project assistant brings state, alarms and relevant documents into the question context.
- 03User outcome
Show forecasts and unusual observations with an explanation; prepare a proposed dashboard or rule for review.
Engineers move from noticing a symptom to investigating it and designing a response.
Site-data mapping and trained-model selection are established during deployment.
SERVICES & EXECUTION ENGINES
Behind each capability, a defined role.
Explore analytical backends, knowledge services, product assistants and edge runtimes through their inputs, outputs and technology. AI engines and integration bridges are described by their actual roles.
Analysis & models
ARG AI BackendData analysis & model lifecycle
A separate numerical-analysis service with language-model provider adapters for industrial and building applications.
- INPUT
- Timestamped data points, device identity and analysis type
- OUTPUT
- Analysis, forecasts, model metrics and job status
Mechanism & product connection
Asynchronous training, versioned model records, prediction storage, metrics and drift monitoring form the lifecycle. Algorithms and validation vary by module; some paths have statistical fallbacks.
Continue to the productProducts & assistants
ArgMainBackEnd / Project AgentOperational data & project assistant
Connects ENID/ACV identity, history, alarms, rules and assistant tools within project context.
- INPUT
- Device messages, user request, project and profile
- OUTPUT
- State, explanations and proposed structure, dashboard or logic changes
Mechanism & product connection
The project assistant is separate from the ML service. Its documented architecture covers equipment/project knowledge retrieval and confirmation of proposed changes; tool and access contracts depend on the deployed version.
Continue to the productKnowledge & operations
ArgDevCatDevice-model & protocol knowledge
Manufacturer, model and protocol mappings support engineering meaning and equipment onboarding.
- INPUT
- Models, technical specifications and point mappings
- OUTPUT
- Device definitions and project engineering context
Mechanism & product connection
The catalog is a knowledge and configuration source. Revisions, units, point mappings and the distinction between general and site settings matter to correct data use.
Continue to the productKnowledge & operations
EKOS Knowledge APIDocument & source-code retrieval
Makes enterprise knowledge searchable with system, repository and evidence provenance.
- INPUT
- Documents, code chunks, queries and source scope
- OUTPUT
- Ranked evidence with source references and identity
Mechanism & product connection
Lexical and semantic retrieval combine through reciprocal rank fusion. Chunk hashes reuse unchanged vectors; model and vector dimension are controlled when building index snapshots.
Continue to the productKnowledge & operations
ARG OperatorEngineering work with explicit tools
Connects knowledge context to work orders, scoped tools, tests and change review.
- INPUT
- Engineering request, prepared context and work scope
- OUTPUT
- Candidates, test results, events and checkpoints
Mechanism & product connection
Candidate preparation, activation and restart are separate phases. Baseline/candidate/live hash comparisons help preserve concurrent changes. Work results and history are retained for reuse.
Continue to the productProducts & assistants
Novexa BackendContext, memory & domain assistants
Combines text, voice and image requests with UDS, memory, domain knowledge and connected services.
- INPUT
- Messages, history and relevant account data slices
- OUTPUT
- Answers, plans, structured fields and tool results
Mechanism & product connection
The modular backend uses routers, domain tables and background workers. Context selection depends on assistant role; ERP, IoT and PC bridges are explicit, configurable connections.
Continue to the productProducts & assistants
JARVIS / Ibn SinaPersian voice interaction
Provides speech-to-text, reasoning/tool use and text/speech responses for the ecosystem.
- INPUT
- Audio or text and conversation session
- OUTPUT
- Transcription, response and synthesized audio
Mechanism & product connection
The desktop version supports local wake words and push-to-talk; the server backend uses configured audio service APIs. Its tool registry covers system, infrastructure and knowledge domains.
Continue to the productAnalysis & models
ArgVisionBackendVisual detection, tracking & counting
Turns camera frames into object detections, tracks and consumable events.
- INPUT
- Camera frames, model configuration and boundaries
- OUTPUT
- Detections, tracks, crossing direction and camera events
Mechanism & product connection
The pipeline separates frame acquisition, detection, tracking and line crossing. A face-matching engine also uses embeddings and cosine similarity; models, permissions and use policies are selected per project.
Continue to the productAnalysis & models
Market Regime BackendMarket state & transition forecasting
Connects market features to semantic regimes, risk context and state-transition forecasts.
- INPUT
- Market observations, features and time contracts
- OUTPUT
- Regimes, destination forecasts and episode ledger
Mechanism & product connection
Current-state classification and next-state prediction are separate tasks. Versioned artifacts, scaling, feature order and semantic mappings must remain compatible with runtime.
Continue to the productAnalysis & models
Price Exhaustion BackendEvent probability & outcome evaluation
Links active market legs, direction-normalized features and time-fixed labels to future-event probabilities.
- INPUT
- Candles, leg features and event-time reference
- OUTPUT
- Calibrated probability, linear evidence and outcome ledger
Mechanism & product connection
Model probability and explanatory score are different outputs. Walk-forward evaluation, model versions and first-bar-state recording preserve evaluation against the original forecast.
Continue to the productEdge & connections
AutoTrade / Execution ServicesBridge from analysis to execution
Connects model analysis to strategies, risk policies and configured Expert Advisor execution paths.
- INPUT
- Market state, signals and risk limits
- OUTPUT
- Trade requests, execution state and feedback
Mechanism & product connection
Forecast records, model outcomes and trade outcomes have distinct responsibilities. Licensing and notifications are supporting services; analytical models operate alongside execution rules.
Continue to the productProducts & assistants
ARG ERP BackendEnterprise context & tools
Exposes documents, reports and business operations to assistant experiences within tenant and user-access context.
- INPUT
- Question, organization identity and permitted records
- OUTPUT
- Reports, contextual answers and domain-tool results
Mechanism & product connection
Enterprise data and capability-scoped tools connect to domain transactions. In the Novexa bridge, inventory, sales and customer context is selected for the relevant query.
Continue to the productProducts & assistants
Agro / Assistant IntegrationCrop knowledge & application context
Connects domain interfaces, crop knowledge and recommendation history to the assistant/service contract.
- INPUT
- Profile, crop context and operator question
- OUTPUT
- Domain recommendations and session history
Mechanism & product connection
This is an application integration. Crop-economics calculations and control rules are distinct from language-model responses; backend and equipment connections are defined during deployment.
Continue to the productEdge & connections
Android ARG AIOn-device language inference
A compatible loaded Android model answers text requests and exposes a local API.
- INPUT
- Prompt and compatible model file
- OUTPUT
- Local model response through JSON
Mechanism & product connection
LocalAiEngine uses MediaPipe, with Ktor routing chat requests to it. Memory capacity, model format and local-network conditions must be evaluated on the target hardware.
Continue to the productEdge & connections
Super AI PLC RuntimeStatistical learning on the MCU
Turns an adaptive signal baseline into an output consumable by BASIC logic.
- INPUT
- Signal samples and learning parameters
- OUTPUT
- Detection variable and persistable model state
Mechanism & product connection
Warm-up, consecutive-event confirmation and save/load/freeze commands define model lifecycle. The runtime cooperates with the main control loop.
Continue to the productEdge & connections
RobotHub / Device BridgesDeliver requests to devices
In Scout, interface commands connect to firmware and telemetry return paths; Novexa’s paired PC bridge also returns operation results.
- INPUT
- Structured command and registered target
- OUTPUT
- Correlated response, state and device results
Mechanism & product connection
These are execution bridges. Agent connectivity does not imply unrestricted device autonomy; each connection defines supported commands and feedback.
Continue to the productShared identity enables precise integration: ENID/ACV for equipment, project identity in the twin, tenants in ERP and UDS in Novexa preserve data context. Connections between services need to carry that context and scope.
THE LEARNING LIFECYCLE
The model is one stage of the journey.
Training, inference and validation must remain connected. Select a stage to inspect its contract and output. Implementation details differ across industrial, trading and controller engines.
DATA
Data & input contract
Define the asset, time range, units, sample quality and feature order. Markets require UTC/timeframe alignment; industrial analysis starts from history and point identity.
Reproducible data and a defined problemSIGNAL & BASELINECONCEPTUAL VIEW
Consumption & comfort forecasting
ARG AI consumes history windows and time features. Without a trained model, energy forecasting can use statistical smoothing; the output’s origin must remain explicit.
Signal anomaly detection
Model-based anomalies are compared with statistical deviations. Meaningful alerts depend on data quality, equipment context and threshold configuration.
Market regimes & transitions
Feature representation, clustering and destination prediction have separate responsibilities. A codebook connects statistical outputs to strategy-consumable meaning.
Calibrated event probability
Price Exhaustion evaluates probabilistic output against time-defined outcomes. Linear-model evidence is a complementary explanation, distinct from probability.
Consumption patterns & lightweight learning
ARG AI uses K-Means for consumption clustering; the controller uses EWMA in ANOM. The method is selected for the data and local execution capacity.
Model quality over time
Error metrics, prediction counts, latency and distribution changes support model review. Monitoring criteria must be defined for the model and project use case.
ARG AI ANALYSIS DOMAINS
Energy & peak demand · Comfort & occupancy · Equipment condition & maintenance · Water & lighting · Datacenter workload, cooling, UPS & capacity
KNOWLEDGE THAT CAN BE USED
Answers with context. Actions with a path.
RAG brings relevant knowledge into answers; agents connect tools to requests. In ARG, these appear in EKOS, the project assistant, Operator and Novexa with distinct roles.
RETRIEVAL AUGMENTED GENERATION
From identified sources to reviewable answers
- 01
Source & identity
Identify the document, repository, version and system; provenance travels with each knowledge chunk.
- 02
Chunking & embedding
Split text into retrievable chunks. Content hashes and model metadata support vector reuse and compatibility.
- 03
Hybrid retrieval
EKOS combines lexical and semantic results with RRF; per-document caps help retain result diversity.
- 04
Evidence-grounded response
Selected evidence and the question enter model context; source identities and references make answers reviewable.
- 05
Review & refresh
Document changes and embedding-model changes are managed separately. Insufficient evidence must remain visible in the user experience.
CONTEXT → TOOL → RESULT
How does an assistant turn a response into work?
- 01
Understand the request & select context
- 02
Select a permitted tool & explicit parameters
- 03
Review or confirmation within the product workflow
- 04
Execute & receive the actual service result
- 05
Record the result & return it to the user
A digital-twin example
“Propose a monitoring dashboard for this asset.” The assistant receives equipment and point context, proposes the appropriate tool and structure, and the confirmed change becomes visible in the project.
EKOS
Knowledge with provenance
Documents, repositories and associated systems remain identifiable in retrieval results.
NOVEXA
The right user context
Memory and relevant UDS slices are selected for the assistant role and question.
OPERATOR
Work with history & recovery
Work orders, tests, candidates and checkpoints make engineering results reviewable.
SEE. HEAR. UNDERSTAND.
The system takes in more than text.
Images, audio and signals each follow a suitable processing path; their structured outputs can feed user experiences and operational workflows.
Vision: images into events
ArgVision connects frame acquisition, ONNX object detection, tracking and directional counting. Face matching uses detection, embeddings and similarity. Novexa Vision/OCR turns document images into reviewable fields.
Voice: conversation into a defined request
JARVIS provides STT, request processing and TTS; desktop wake words use Vosk. Novexa resolves speech text into intent and structured parameters, selects a registered tool and returns its service result.
ENGINEERED AS A SYSTEM
Technology selected for a specific role.
The architecture combines numerical engines, language models, knowledge services, operational tools and edge runtimes. Each technology family has a defined place.
Numerical models
Java industrial engines and Python analytical paths in AutoTrade; model selection depends on the task and data.
Language models & providers
ARG AI has multiple provider adapters. Model choice, API versus local execution and service capacity are configured per deployment.
Knowledge & retrieval
EKOS connects text, vectors and retrieval snapshots to source identities.
Vision & speech
ArgVision, JARVIS and Novexa multimodal workflows provide distinct experiences with explicit inputs and outputs.
Backend & data
Stacks differ by service: Novexa uses a Python HTTP server; ERP uses FastAPI.
Edge & communication
From Android local inference to statistical controller learning and robot command bridges, each environment has its own contracts and capacity.
FROM ARCHITECTURE TO OPERATION
Intelligence where the problem requires it.
Execution location depends on data, latency, hardware capacity and connectivity. A project can combine several paths.
In the organization
Industrial data ingestion, ARG AI, enterprise knowledge and optionally Ollama operate within project infrastructure.
- Asset / tag / ENID and time mapping
- Knowledge sources, access and model versions
- Capacity testing and use-case acceptance
Build a measurable pilot
Problem
A specific question and expected outcome
Data
Sources, quality, time and ownership
Measure
Error, answer quality or tool outcome
Acceptance
Evaluation with project data and users
From reading to implementation design
CLEAR QUESTIONS. PRECISE ANSWERS.
Key questions
Useful distinctions when selecting and designing a use case.
How do ML, LLMs and agents differ here?
ML learns patterns and predictions from numerical data. LLMs transform language and evidence into explanations or structures. Agents coordinate tool selection and execution in a workflow. The three can cooperate while retaining distinct inputs, measures and responsibilities.
Which parts require actual site data?
Consumption forecasts, anomalies and equipment condition require relevant validated data. Knowledge assistants require approved sources; operational tools require service connections and permissions. Inputs, expected outputs and tests are defined per use case at project start.
Is RAG the same as training a language model?
RAG retrieves relevant knowledge at request time and supplies it as answer context. Language-model training is a different process. This architecture distinguishes numerical-model training from knowledge retrieval and language-model consumption.
How do we know a command actually executed?
A text response, accepted request, delivered command and observed device change are different stages. Each product’s execution path must return its required tool result, receipt or observed state; industrial projects define and test this contract during site engineering.
How does an AI project start?
Choose a measurable problem, such as one asset’s consumption forecast or questions over a defined document set. Specify data sources, use-case ownership, expected outputs and acceptance criteria, then evaluate a pilot with that environment’s data and users.
BUILD YOUR NEXT INTELLIGENT SYSTEM
Your problem is where intelligence begins.
Choose an asset, a knowledge collection or a workflow. Let’s design the path from data to outcome around that use case.