Self improving risk data agent handler

Let me look at the project to write an accurate writeup.

Based on your project (Audit Passport — the five-agent Data Rescue pipeline on Cognee), here's a concise Devpost writeup:

Inspiration

A manufacturing compliance officer discovers her records are corrupted four days before a regulatory audit — duplicates, unit conflicts, contradictory numbers, broken evidence chains. She can't write SQL and has no time to hand-check thousands of rows. We wanted an agent that doesn't just clean data, but finds, fixes, and explains it in plain language — and gets smarter every time new evidence arrives.

What it does

Audit Passport ingests real manufacturing data and runs five collaborating agents that find broken records, rank them by audit risk, decide how to fix them, and write a downloadable auditor-ready report. Every finding shows its source → row → field → conflicting value, a reason for why it's broken, the recommended fix tier, and a PyMC-derived audit-blocker probability with a credible interval. The first screen is a risk dashboard — readiness score, urgent issues, fixable-now count — not a JSON dump.

How we built it

• Five-agent pipeline (agents.py): Find It → Rank It → Act On It → Explain It → Show It.
• Cognee as the shared memory layer — each agent reads prior agents' cards and writes its own back, so the handoff is inspectable and the system remembers instead of re-running.
• Deterministic detectors surface candidates (duplicates, unit conflicts, impossible dates, orphaned references…); agents reason over them — no hardcoded findings.
• PyMC for probabilistic risk scoring, Geodo for real-world domain research woven into the narrative, and a Flask /api/run web server driving the dashboard and PDF export.

Challenges we ran into

• Making agents actually collaborate through memory rather than each re-deriving everything — getting the Cognee session handoff right took the most iteration.
• Normalizing severity and action tiers so the audit metrics stayed correct across agents.
• Keeping everything explainable end-to-end — "the model said so" had to appear nowhere.

Accomplishments that we're proud of

• A judge can run it cold from the dashboard to a downloadable report with no coaching.
• Seven detected problem classes, exceeding the minimum.
• The Memory Ripple moment: adding new evidence writes a patch to Cognee and the affected finding, its risk score, action, and PDF all update from shared memory.

What we learned

Memory is what turns a chain of prompts into a system. Once agents share state through Cognee, "self-improving" stops being a slogan — new evidence genuinely changes downstream decisions without a full re-run. We also learned how much explainability depends on a strict card schema enforced from the start.

What's next for Self Improving Risk Data Agent Handler

• A fuller memory-governance / LINT layer so the system flags stale or conflicting memories on its own.
• Continuous learning from auditor feedback to tune ranking over time.
• More data connectors and live ingestion so the passport stays current between audits.

Want me to drop this into a file (e.g. DEVPOST.md)?

Built With

• codex