Audit Rescue Desk

Inspiration

Manufacturing compliance teams often face audit pressure with messy operational data: duplicate records, inconsistent units, impossible timelines, missing values, and customer references that cannot be traced. For a non-technical compliance officer, the problem is not just finding bad rows. The harder problem is explaining what happened, deciding what to fix first, and proving that every decision was handled responsibly.

Audit Rescue Desk was inspired by that moment: a team has corrupted warehouse data four days before a regulatory audit, and they need a workflow that feels understandable, defensible, and human-controlled.

What We Built

Audit Rescue Desk is an agentic audit workflow for manufacturing data rescue.

The product helps users:

• Load or upload manufacturing CSV datasets
• Preview raw data without writing SQL
• Detect data quality issues
• Group findings into audit-relevant problem types
• Recommend risk priority and remediation actions
• Let humans override risk priority and final decisions
• Store agent handoffs and user policy decisions in Cognee memory
• Export a plain-English audit report

The long-term product idea is Policy Memory: the system should learn how each compliance team ranks risk and handles remediation over time. The first audit finds the problems. The next audit starts with the team’s remembered policy.

How It Works

We designed the workflow around specialist agents:

1. Data Forensics Lead detects corrupted records and records row-level evidence.
2. Audit Risk Triage recommends risk level and P0-P3 fix priority.
3. Compliance Action Advisor suggests safe human-approved remediation.
4. Audit Narrative Writer turns the decision trail into an audit-ready report.
5. Policy Learner is the future-facing memory layer that summarizes user risk choices and remediation preferences.

Cognee is used as the memory layer. The product is designed to write agent handoffs, user risk overrides, remediation choices, and natural-language policy notes into Cognee memory, then recall those policies in future reviews.

Geodo supports the domain research layer: customers, companies, market context, acquired plants, traceability, and audit defensibility.

The Kaggle Track 01 Harven Manufacturing dataset gives the product a realistic benchmark scenario.

How We Built It

We built the app with:

• React + Vite + TypeScript
• Papa Parse for CSV parsing
• Local deterministic audit rules for stable, explainable findings
• A local Cognee SDK bridge for memory operations
• Markdown-based agent playbooks so users can customize agent roles
• Downloadable Markdown audit reports
• A dashboard-first UI for non-technical compliance users

We intentionally kept the data detection layer deterministic. In compliance workflows, explainability matters more than opaque automation. The system should not silently overwrite data or invent findings. Every issue needs evidence, reason, priority, and a human decision.

Challenges

The biggest challenge was balancing agentic behavior with audit trust.

A fully autonomous agent could feel impressive, but compliance users need control. We had to design agents as bounded specialists with visible reasoning rather than black-box decision makers.

Another challenge was product structure. Early versions showed too many cards and internal hackathon details. We redesigned the experience into three layers:

• Dashboard for daily audit work
• Design Agents for configuring the agent playbook
• Tutorial for product narrative and workflow explanation

We also had to think carefully about Cognee usage. Instead of sending a full 5,000-row CSV into memory, we designed a low-cost memory strategy: store concise summaries, agent handoffs, user policy notes, and final remediation choices.

What We Learned

We learned that data rescue is not only a technical problem. It is a workflow problem.

A useful audit product must answer:

• What is wrong with the data?
• Why does it matter?
• What should be fixed first?
• Who approved the decision?
• What did we learn for next time?

We also learned that memory is a powerful product differentiator. Cognee is not just a place to store logs. It can become a policy layer that helps the product adapt to each compliance team’s judgment over time.

What’s Next

Next, we want to:

• Complete category-first data quality review
• Add final remediation plan selection per issue category
• Use Cognee recall to influence future risk ranking
• Use Cognee improve for long-term policy memory
• Add real Geodo research artifacts into the product
• Add optional LLM-generated audit narratives while keeping findings deterministic
• Export cleaned draft CSVs with human-approved changes only

Audit Rescue Desk aims to become a learning audit operations layer for manufacturing compliance teams.

Built With

• codex
• cognee
• geodo
• openai
• python
• trupeer