🛡️ AuditGuard AI

Multi-Agent Data Rescue & Regulatory Audit System

4 AI agents that scan corrupted manufacturing data, rank every issue by audit risk, fix what they can, and deliver a regulator-ready PDF narrative — all in under a minute.

🏆 Track 01 — Data Rescue · M-AGENTS Hackathon · June 2026

---

The Problem

Manufacturers face FDA audits with thousands of rows of production data that contain duplicates, unit conflicts, outliers, and compliance contradictions. Catching these manually takes days. Missing them costs millions.

The Solution

AuditGuard AI is a 4-agent pipeline where each agent reads from and writes to a shared Cognee memory layer — real handoffs, not file passing:

CSV Upload
    │
    ▼
Agent 1: Scout ──── detects issues ────► Cognee Memory
                                               │
Agent 2: Ranker ◄── reads findings ───────────┤
    │ ranks by audit risk ────────────────────►│
                                               │
Agent 3: Fixer ◄──── reads ranked list ───────┤
    │ auto-repairs what it can ───────────────►│
                                               │
Agent 4: Narrator ◄── reads ALL memory ───────┘
    │
    ▼
📄 Regulator-Ready PDF Download

---

What It Detects

Issue Type	Example
Exact duplicates	Same row appears 8 times
Near-duplicates	Same lot, quantity off by 1–2
Unit conflicts	kg vs lbs in same column
Statistical outliers	Temperature 3σ from mean
Missing timestamps	batch_date is null
Compliance contradictions	PASS status + temp > 85°C (FDA auto-fail)

Regulatory standards cited: FDA 21 CFR Part 11 and ISO 13485

---

Tech Stack

Component	Technology
Multi-agent memory	Cognee (with in-memory fallback)
LLM reasoning	Anthropic Claude
Backend	FastAPI + Python
Frontend	React (CDN, no build step)
Output	PDF with signature block

---

Quickstart

# 1. Install dependencies
cd backend
pip install -r requirements.txt
cp .env.example .env

Add to .env:

ANTHROPIC_API_KEY=your_key
COGNEE_API_KEY=your_key   # free 14-day trial at cognee.ai

No Cognee key? The pipeline automatically falls back to in-memory store — it still runs fully.

# 2. Start backend
uvicorn main:app --reload --port 8000

# 3. Open frontend
open frontend/index.html   # No build step — uses CDN React

Run the demo

1. Drag data/sample.csv into the upload zone
2. Watch all 4 agents fire in sequence with real-time status updates
3. Download the signed-ready audit narrative PDF

---

Why the Judges Liked It

Criterion	How AuditGuard AI delivers
Data rescue	Detects 6 issue types across 200 rows automatically
Multi-agent architecture	4 agents with real Cognee memory handoffs
Explainability	Every decision has a specific plain-English reason — "The model said so" never appears
Non-technical usability	Drag-and-drop UI → PDF with signature block, no engineer required
Regulatory relevance	FDA 21 CFR Part 11 + ISO 13485 cited; compliance contradictions auto-escalated

---

Agent Responsibilities

Scout — scans every row, classifies issue type and severity, writes findings to Cognee

Ranker — reads Scout's findings, scores by audit risk using Claude + FDA context, writes prioritized list

Fixer — reads ranked issues, auto-repairs what's safe to fix (dedup, unit normalization), logs actions

Narrator — reads all memory, generates plain-English audit narrative with regulatory citations, exports PDF

---

Built with Claude + Cognee · M-AGENTS Hackathon 2026