Deploy the Agentic Workflow for the Multimodal Weld Defect Detection Sample Application#

This section shows how to deploy the multimodal sample application with the agentic workflow enabled. The multimodal application produces fusion results, and when new fusion results arrive, the meta-agent in the agentic stack is triggered. The meta-agent, powered by an LLM served through OpenVINO™ model server, produces structured policy decisions, root-cause analysis, and maintenance tickets.

Architecture Overview#

The agentic workflow is implemented as a LangGraph framework-based, sequential multi-agent pipeline. The apm-agent, which is the meta-agent, acts as the orchestrator that triggers the workflow when new fusion results arrive and coordinates the execution of specialized agents, each responsible for a distinct stage of reasoning. Each agent consumes the shared execution context together with outputs from previous stages and produces traceable intermediate artifacts and a final maintenance recommendation.

Agentic Workflow Architecture Diagram

Agent

Input

Output

Policy Agent

Fusion results (fusion_result) filtered according to the configured analysis thresholds. Uses fused_decision, fusion_confidence, modality confidence scores, anomaly indicators, and timestamp alignment (vision_rtsp_ts_diff_ms).

Structured policy violation report containing the detected defect class, fusion and modality confidence scores, alignment quality, and preliminary priority. Policy decisions are based on fusion_mode and configured severity rules.

Analysis Agent

Policy output as the primary anchor, supplemented by fusion, vision, and time-series classifications for the detection window when available.

Policy-anchored analysis summarizing the policy finding and corroborating it with modality classification data and confidence evidence; falls back to event-level or summary-level analysis when no policy output is available.

Ticketing Agent

Policy evaluation and root-cause analysis results.

Structured maintenance ticket containing the priority, title, description, affected component (if available), recommended action, estimated resolution time, and defect class tags. Ticket priority and escalation follow the configured ticketing rules.

Note

The [SYSTEM] prompt provides shared domain knowledge, including the canonical defect taxonomy, label normalization rules, and available fusion data. It establishes the common reasoning context for all agents and is not a separate execution stage.

Minimum System Requirements#

In addition to the minimum requirements and software requirements of the sample application, the agentic workflow requires the following:

Component

Minimum Requirement

Hardware

Intel® Core™ Ultra Series 3 (Panther Lake/PTL) processor or newer

RAM

Minimum 48 GB (recommended 64GB)

Storage

Minimum 512 GB

The platform check accepts Panther Lake (PTL) and newer Intel Core Ultra series, including Nova Lake (NVL), the codename for Intel’s upcoming Core Ultra Series 4 processors. It checks the Intel vendor and Core Ultra series in the CPU model name. For unbranded engineering samples (for example, Genuine Intel(R) 0000), it also recognizes PTL by CPU family 6/model 204 and NVL by family 18/model 1 or 3. Unknown unbranded platforms are rejected rather than inferred from higher CPU family/model numbers. Run make check_hardware to check the CPU before deployment.

Passing this check does not imply platform validation. Newer platforms also require compatible GPU drivers and inference runtimes; only the hardware listed below has been validated.

Validated Platforms#

Product / Family

Validated inference devices

Intel® Core™ Ultra Processors Series 3 (Intel Core Ultra 7 356H, 16 Cores, 48GB)

CPU, iGPU

Prerequisites#

  1. Ensure the .env file is configured with valid values for:

    Note

    • If you are trying to deploy it on Ubuntu 26.04, update the DLSTREAMER_PIPELINE_SERVER_IMAGE value in .env to intel/dlstreamer-pipeline-server:2026.3.0-ubuntu26-rc1 before deployment.

    • HOST_IP

    • INFLUXDB_USERNAME, INFLUXDB_PASSWORD

    • VISUALIZER_GRAFANA_USER, VISUALIZER_GRAFANA_PASSWORD

    • MTX_WEBRTCICESERVERS2_0_USERNAME, MTX_WEBRTCICESERVERS2_0_PASSWORD

    • S3_STORAGE_USERNAME, S3_STORAGE_PASSWORD

    • SEAWEEDFS_WEB_AUTH_USER, SEAWEEDFS_WEB_AUTH_PASSWORD

    • VLLM_GPU_MEMORY_UTILIZATION: fraction of the memory available to the GPU that vLLM preallocates (VRAM on discrete GPU; shared system memory on integrated GPU). Set it to memory to allocate to vLLM (GB) / total memory available to the GPU (GB). For the default setup (Qwen3.5-2B fp8 + LoRA adapter, 8192 max context), vLLM is budgeted 8.5 GB: set 0.27 for 32 GB, 0.19 for 45 GB, or 0.13 for 64 GB of memory.

  2. Review the licenses of the models prior to deployment. make up_agentic downloads them automatically on the first run:

Deploy the Agentic Workflow#

Run the agentic stack (downloads the LLM model first, then starts the containers):

Note

  • Model download time varies depending on network speed and hardware.

  • The service is polled every 5 seconds for up to 50 minutes.

  • Supported devices for Agentic Workflow are : CPU, GPU

cd industrial-edge-insights-multimodal
make up_agentic

To include the MCP server, set MCP_USERNAME and MCP_PASSWORD in .env, then run:

make up_agentic MCP=true

See Use the MCP server for the tools and the example client.

For a fresh build before deployment:

cd industrial-edge-insights-multimodal
make build
make up_agentic

Running the Agentic Workflow on GPU#

By default, the agentic workflow is configured to run on CPU.

To trigger the agentic workflow on GPU, update LLM_DEVICE in .env to GPU:

vi .env
# change LLM_DEVICE to GPU
LLM_DEVICE=GPU
# Deploy Agentic Workflow
make up_agentic

Configure the Agent#

Use-Case Configuration#

The agent behavior is controlled by configs/agentic/agents.yaml:

Field

Default

Description

analysis.detection_schema

fusion_result

Instructs the agent to read fusion_result records rather than raw bounding-box detections

analysis.min_confidence

0.5

Minimum fusion_confidence to include a record in the analysis window

analysis.max_detections_per_run

1000

Maximum number of fusion_result records fetched per run

policy.alert_threshold

0.75

fusion_confidence threshold to raise a policy violation

policy.fusion_mode

OR

AND requires both modalities to be anomalous; OR triggers on either

policy.critical_classes

Burnthrough, Lack of Fusion, and Crater Cracks

Classes that produce CRITICAL severity regardless of threshold

ticketing.backend

jira

Ticket destination: jira, servicenow, or none

ticketing.auto_create

true

Submit tickets automatically after each completed run

Defect Class Taxonomy#

The policy and ticketing agents operate on the following class hierarchy:

Priority

Classes

CRITICAL

Burnthrough, Lack of Fusion, and Crater Cracks

HIGH

Excessive Penetration

MEDIUM

Porosity, Porosity with Excessive Penetration, Undercut, and Warping

LOW

Overlap, Spatter, and Excessive Convexity

Non-actionable

Good Weld, No Weld, and No Label

Prompts#

Agent reasoning prompts are in configs/agentic/prompts/weld-quality-monitoring.txt. Each [TAG] block maps directly to an agent stage:

Section

Controls

[SYSTEM]

Canonical class labels, label normalization rules (No_Weld → No Weld, Good_Weld → Good Weld, and Porosity_w_Excessive_Penetration → Porosity with Excessive Penetration), and available fusion fields

[POLICY]

How violations are identified: fusion_confidence as the primary signal, priority determined by defect class, confidence used only to verify the configured threshold is met, and fused_decision reported exactly when present

[ANALYSIS]

Policy-anchored analysis treating the policy decision as the source of truth; corroborates with fusion and modality data when available; falls back to event-level or summary-level analysis when no policy output exists

[TICKETING]

Escalation rules tied to defect class and fusion_confidence threshold: CRITICAL ≥ 0.80, HIGH ≥ 0.75, MEDIUM ≥ 0.65, LOW ≥ 0.55; ticket generated only when fused_decision == 1

Fallback Policy#

configs/agentic/policy_fallback.json defines per-class thresholds and actions used by apm-agent when LLM_MODE=fallback. Available actions:

Action

Description

HALT_LINE

Stop the production line immediately

REDUCE_HEAT_INPUT

Reduce welding current or power

SCHEDULE_INSPECTION

Flag for next-shift inspection

ADJUST_PARAMETERS

Adjust process parameters

CHECK_FIXTURING

Check part fixturing and alignment

MONITOR

Continue monitoring without action

CONTINUE

No action required

Verify the Deployment#

  1. Check overall stack health:

    cd industrial-edge-insights-multimodal
    make status
    
  2. Check the output in Grafana dashboard:

    • Use the link https://127.0.0.1:3000 to open Grafana dashboard in a browser, preferably the Chrome browser.

    • Log in to Grafana dashboard using the VISUALIZER_GRAFANA_USER and VISUALIZER_GRAFANA_PASSWORD values from the .env file:

      Grafana dashboard login

    • After logging in, click Dashboards and then select Multimodal Weld Defect Detection - Agentic Dashboard: Menu view

    • The following pages appear: Multimodal Weld Defect Detection Agentic Dashboard

      Results of the multi-agent pipeline appear: Agentic Results for weld data

      The detections dashboard appears: Detections Results for weld data

Stop the Stack#

cd industrial-edge-insights-multimodal
make down