AssedGuard AI

Inspiration

AssedGuard AI was inspired by a realistic manufacturing audit crisis: a company has an FDA-style audit coming soon, but its production data cannot be trusted.

In our result report, AssedGuard AI scanned 5,000 manufacturing records from track01_data_rescue.csv and identified 1,098 data integrity issues. Out of these, 1,093 were critical, 5 were moderate, and 0 were minor. That immediately showed us that data quality is not just a technical issue — it can become a serious compliance, traceability, and business risk.

We built AssedGuard AI for a user like Sarah, a non-technical compliance officer who does not know SQL or Python but still needs to understand what went wrong, which records are risky, and what actions must be taken before regulatory submission.

Our goal was to turn messy manufacturing data into a clear, explainable, audit-ready action plan.

What it does

AssedGuard AI is a five-agent data rescue system for manufacturing compliance teams.

A user uploads a manufacturing CSV file, clicks Run Audit, and the system analyzes the dataset through five coordinated agents:

Scout Agent detects data integrity issues such as impossible values, decimal-shift errors, out-of-range sensor readings, and corrupted production records.
Ranker Agent prioritizes each issue by severity, audit impact, quality impact, and traceability risk.
Fixer Agent recommends safe corrections or review actions. In the report, 5 issues were auto-corrected with logged reasons.
Counsel Agent reviews the actions for compliance safety and decides which findings need human sign-off.
Narrator Agent generates the final audit narrative and pre-audit action plan.

The system does not simply say “errors found.” It explains what happened, why it matters, what action was taken, and what still requires human verification.

For example, the report found physically impossible negative quantity values, decimal-shift errors that could corrupt production totals, and moderate out-of-range sensor readings that may indicate sensor faults.

A simple version of our risk logic is:

[ RiskScore = Severity + AuditImpact + QualityImpact + TraceabilityImpact ]

This helps the system separate routine cleanup from issues that could create audit failure risk.

How we built it

We built AssedGuard AI as a structured multi-agent pipeline.

The workflow starts when the user uploads the manufacturing CSV file. The system then scans the records, detects data integrity problems, ranks them by risk, recommends corrective action, checks compliance safety, and generates a downloadable audit narrative.

Our agent pipeline is:

[ Scout \rightarrow Ranker \rightarrow Fixer \rightarrow Counsel \rightarrow Narrator ]

The Scout Agent identifies the raw data problems. The Ranker Agent turns those findings into prioritized risks. The Fixer Agent suggests or applies safe corrections when appropriate. The Counsel Agent adds a compliance review layer so high-risk findings are not blindly changed. The Narrator Agent turns the complete investigation into a readable audit document.

We used Cognee as the shared memory layer so the agents could work together using structured knowledge rather than isolated prompts. We used Trupeer for the demo recording and Geodo as part of the regulatory context layer. In the generated report, Geodo regulatory context was included as a simulated profile, showing overall audit exposure as low while still preserving the audit review workflow.

The final output is a confidential audit document with an executive summary, findings and actions taken, open items requiring human sign-off, regulatory context, and a pre-audit action plan.

Challenges we ran into

One challenge was making the agents work as a connected system instead of five separate AI responses. To solve this, we designed each agent with a specific responsibility and structured handoff.

Another major challenge was correction safety. In manufacturing compliance, automatically changing records can be risky. If a value is physically impossible, the system should not simply guess the correct value. In our report, impossible quantity values were escalated for source verification instead of being blindly corrected.

We also had to handle severity carefully. The report contained a very large number of critical findings, especially decimal-shift issues in quantity and weight fields. These are dangerous because they can corrupt production totals and misrepresent batch records.

A final challenge was making the result useful for a non-technical user. A compliance officer should not need to inspect raw tables line by line. The system needed to summarize the findings, explain the risk, and produce an action plan that could be understood before an audit.

Accomplishments that we're proud of

We are proud that AssedGuard AI produced a complete audit-style report from a corrupted manufacturing dataset.

The system scanned 5,000 records and identified 1,098 data integrity issues. It separated the findings into critical and moderate risks, logged correction reasons, and escalated high-risk records for mandatory human sign-off.

We are especially proud of the five-agent design. Each agent has a clear role:

Scout finds the problem.
Ranker decides how serious it is.
Fixer recommends or applies safe action.
Counsel checks compliance safety.
Narrator creates the final audit story.

We are also proud that the final result is not just a dashboard. It is an audit narrative with traceable reasons, open items, sign-off sections, and a pre-audit action plan.

That makes AssedGuard AI feel closer to a real compliance workflow rather than a basic data-cleaning tool.

What we learned

We learned that data quality in manufacturing is much bigger than missing values or duplicate rows.

Bad production data can affect batch traceability, quality decisions, regulatory confidence, and audit readiness. A decimal-shift error in quantity or weight may look small in one row, but across production records it can corrupt totals and create serious compliance risk.

We also learned that multi-agent systems are stronger when each agent has a narrow, clear job. Instead of asking one model to do everything, we split the workflow into detection, ranking, correction, compliance review, and narration.

Most importantly, we learned that compliance AI should keep humans in control. The safest system is not the one that automatically changes every suspicious record. The safest system is the one that explains the issue, logs the reason, recommends action, and escalates uncertain or high-risk findings for human verification.

What's next for AssedGuard AI

Next, we want to make AssedGuard AI more production-ready for real manufacturing teams.

Future improvements include:

better support for very large manufacturing datasets,
plant-by-plant issue comparison,
stronger root-cause analysis,
approval workflows for human sign-off,
deeper integration with quality management systems,
live Geodo regulatory checks,
better visual dashboards for critical and moderate findings,
correction history with before-and-after values,
and customizable compliance rules for different industries.

We also want to improve the Counsel Agent so it can generate stronger audit-readiness guidance, identify which issues require mandatory review, and help teams prepare documentation before regulatory submission.

Our long-term vision is for AssedGuard AI to become a trusted audit-preparation assistant for manufacturing teams.

It helps teams catch corrupted data early, explain every decision clearly, and walk into audits with confidence.