This project contains implementation of a multi agent ai system which can classifies emails, json and pdfs. It is able to extract intents, tone and risk associated with the documents mentioned before.
Detects format and business intent
Classifies: JSON, Email, PDF
Extracts Intent: RFQ, Complaint, Invoice, Regulation, Fraud Risk
Passes routing + intent metadata to memory
Extracts structured fields: sender, urgency, issue/request
Identifies tone (e.g., escalation, polite, threatening)
Triggers action based on tone + urgency
Escalate → notify CRM (simulate with API)
Routine → log and close
Parses webhook data (simulate with provided sample)
Validates required schema fields
Flags anomalies (e.g., field mismatch, type errors)
If anomalies detected → log alert in memory/API
Extracts fields (using PDF parsers, not hardcoded text)
Parses line-item invoice data or policy document
Flags if:
Invoice total > 10,000
Policy mentions "GDPR", "FDA", etc.
All agents read from and write to a centralized memory store.
Stored Data Includes:
- Input metadata (source, timestamp, classification)
- Extracted fields per agent
- Chained actions that were triggered
- Agent decision traces
Triggers follow-up actions based on agent outputs.
Examples:
- Create a support ticket
- Escalate an issue
- Flag a compliance risk
These actions can be simulated via REST calls:
POST /crm
POST /risk_alert-
User uploads email → Classifier detects: Email + Complaint
-
Email agent processes → Tone = angry, Urgency = high
-
Action Router calls: POST /crm/escalate
-
Memory logs entire trace for audit
-
-
Input:
Output:
Logs:
-
Input:
Output:
Logs:
-
Input:
Output:
Logs:
-
Input:
Output:
Logs:
-
