Industry applications Manufacturing

Defects found.
What happens next?

Manufacturing AI, grounded in process data and expert judgment.

What changed in the process? Which production was exposed?
Follow the team's investigation criteria and give the factory evidence for its next decision.

Explore the analysis report Read the source case (Korean)

An analysis example from the HEARTCOUNT blog using synthetic SMT data. Approval, control integration, and execution are future directions.

Quality investigation in electronics, EMS, and PCB manufacturing

Six boards failed inspection.
Forty-two shared the conditions.

The BGA solder-joint defect rate has risen. Finding the failed boards is only the start: connect process conditions with production history to investigate possible causes and trace everything that may be affected.

Synthetic data example

Investigating a rise in BGA defects

Equipment: REFLOW-C-01
Example period: Aug 10–25, 2026

The production team's question

Why did the defect rate rise, and which boards need attention?

How the agent investigates

  1. Check the metric

    Compare the BGA defect rate with the operation’s 0.35% ceiling to define the investigation.

  2. Compare process conditions

    Examine V6 and V7 together: temperature setpoints, measured temperatures, and defect rates.

  3. Trace the exposed production

    Trace all 42 boards processed under V7, including the 36 that passed initial inspection.

  4. Prepare the report

    Separate possible causes, affected production, and the questions the factory still needs to resolve.

What the analysis found

The evidence points first to a setpoint change.

V7’s Zone 6 setpoint was 245℃, compared with 232℃ for V6. Measured temperatures were close to the setpoints in both versions, making the unapproved recipe change an earlier investigation priority than equipment failure.

Thirty-six boards passed inspection under the same settings.

Of the 42 boards processed under V7, 36 passed their first inspection. Their shared process conditions make their location, shipping status, and need for reinspection relevant to the review.

A report of findings and proposed reviews from the published synthetic-data analysis. V7’s sample of 42 boards does not establish a cause. Approval, control integration, and execution remain future directions.

Inspect the data and evidence

What changed?

Two versions of the equipment's process recipe, as compared in the original case.

MeasureV6V7
Approval statusApprovedUnapproved
Zone 6 setpoint232℃245℃
Zone 6 measured average231.97℃244.97℃
BGA defect rate0.28%14.29%
Boards processed6,40842

V7 has a small sample of 42 boards. The difference alone does not establish a cause.

How far does the exposure extend?

42boards

Processed under V7
3 work orders · 3 lots

Passed inspection
36boards Review for reinspection
Defects confirmed
6boards 2 scrapped · 4 for rework

The current locations and shipment status of the 36 boards that passed inspection still need checking. Trace all production that shared the conditions, including units with no defect recorded.

What still needs checking

Review the full temperature profile, the history of the recipe change, and shipment records. Possible causes identified in the analysis need validation on the factory floor.

Review the evidence and data limitations

Source: HEARTCOUNT's “Defects found. What happens next?” blog case, based on synthetic SMT data. The source article is in Korean.

Make expert investigation methods available to AI

The data tells you what happened.
Expertise tells you where to look.

A metric definition does not tell an agent how to investigate. The team also needs to define what counts as a problem, what to check first, and which conditions make a fair comparison.

KPI

When should an investigation start?

The case uses a BGA defect-rate upper limit of 0.35%. The agent starts from a threshold defined by the organization.

Driver

What should the agent check first?

Review reflow, printing, and material conditions. Equipment, process recipes, SPI measurements, and material lots define the investigation path.

Lever

What makes a useful comparison?

Narrow the analysis from equipment to work order to lot. Compare against normal production with the same product, revision, line, and component position.

Capture these investigation paths and comparison rules in a Decision Graph. In the published case, the agent used MCP to retrieve data and decision criteria and assemble the analysis.

Explore the role of Decision Layer

Beyond the analytical report

Choose the next step.
Then check what actually changed.

Follow the path from investigating possible causes and affected production to reviewing an action, approving it, requesting execution, and checking the outcome. The example distinguishes the analysis from future integrations.

Illustrative workflow: steps 1–2 reflect the blog's analysis and recommendations; steps 3–6 describe future directions.

01 / 06

Published analysis example

Connect the quality issue to its production conditions.

With metrics and investigation criteria connected, an agent can narrow possible causes and trace the affected production. The manufacturing data model and investigation paths require configuration for each factory.

What the case investigated

The quality issue
BGA defect rate and reflow equipment
Conditions compared
Recipe version, approval status, setpoint, and measured temperature
Production exposed
Work orders, lots, and boards processed under the same conditions

Follow the links from equipment to work orders and lots, keeping the comparisons and affected quantities available for review.

02 / 06

Actions proposed in the analysis

Review actions against the evidence and the production exposed.

Possible follow-ups include stopping V7, restoring an approved recipe, locating and reinspecting boards that passed, and preventing the use of unapproved recipes.

Questions to resolve before execution

Scope
Which equipment, recipe, and production units would the action affect?
Operational impact
Who reviews the quality risk and the impact on production and shipments?
Open questions
Which data and on-site checks are still needed before a decision?

The factory needs to validate and approve any proposed action. This example does not include measured results after an intervention.

03 / 06

Future direction

An authorized person approves the action and its conditions.

Quality and production owners review the proposed scope, evidence, and execution conditions. A change in those conditions should trigger another review.

What an approval should capture

Who
The person authorized for this type of action
What
The equipment and production in scope, permitted changes, and validity period
Under which conditions
Quality requirements, plant status, and conditions for stopping or reviewing the action

Operational approval requires a separate integration. It is distinct from the current framework's ability to ask a person for input during an analysis.

04 / 06

Future direction

Send the approved action to existing operations and control systems.

An MES integration or execution module could request work instructions and approved setting changes. Existing controllers retain responsibility for fast control loops and safety functions.

Responsibilities across the systems

Decision Layer
Manage the evidence, selected action, and constraints
Execution module
Check permissions and approval, then make the system-specific request
Existing systems
Perform permitted operations or equipment control and report their status

MES and equipment-control connectors are not included in the current public framework.

05 / 06

Future direction

Check that the change reached the factory floor.

Sending a command does not mean the action is complete. Distinguish accepted, in progress, completed, and failed requests, then verify the change against the actual plant state.

Evidence of execution

Request record
Target, timestamp, approval, and request identifier
Processing status
Accepted, in progress, completed, or failed
Applied change
Factory records showing the active recipe or work status

Review both the execution system's response and the factory's actual records.

06 / 06

Future direction

Verify the action and its business impact separately.

After confirming the change was applied, compare quality, productivity, and side effects under comparable production conditions. Keep the results as evidence for future decisions.

How to assess the outcome

Execution
Was the approved action applied to the intended target?
Performance
How did defects, rework, throughput, and delivery change?
Future use
Under which conditions could the team use this action again?

Consider automatic execution only for work with verified effects and well-defined conditions. This analysis example contains no measured post-action improvement.

Toward Physical AI

Connect factory-wide judgment
with action on the floor.

Look beyond the performance of a single machine.
Make decisions for the production flow as a whole.

Bring together production, quality, equipment, and material data with shift logs, maintenance records, and operating rules. Evaluate possible actions against throughput, quality, cost, and delivery across the factory.

HEARTCOUNT's direction is to extend from process analysis into existing MES, operations systems, and equipment controllers. The test is whether an action fits the situation and improves actual production performance.

What you can use today.
Where we plan to go next.

Configure analyses for your process data and reuse reviewed procedures. The Always-on Decision Agent and approval and execution interfaces are planned for open-source release.

Available in the public framework

Apply your team's criteria and inspect the evidence

  • Query data through defined metrics
  • Reuse registered analytical code and procedures
  • Inspect inputs, queries, results, and validation records
  • Call analyses from agents and business applications
Review the code and current scope
Requires factory-specific setup

Investigation criteria for your production process

  • Quality and productivity metrics, with agreed exception thresholds
  • Data relationships across processes, equipment, lots, and time
  • Investigation order and comparison criteria
  • Supporting documents, operating history, and data permissions

Adapt and validate the BGA example's data model and decision criteria for your own process.

Planned open-source release

From proactive investigation to connected action

  • Always-on Decision Agent briefings
  • Review and approval by the responsible person
  • Execution requests to operations and control systems
  • Broader automation after verifying execution and performance

The Always-on Decision Agent and approval and execution interfaces are planned for open-source release and are not part of the current public framework. System connectors and approval conditions are configured during implementation.

Start with one quality issue
that takes too long to investigate.

Agree on the performance metric, available data, and the factory's investigation criteria.
Then consider which actions and systems a future integration should support.

Contact us (Korean)Explore other industry applications