Developed by Sabrina Palis
MSc Artificial Intelligence (First Class Honours)
AI Engineering • Multi-Agent Systems • Workflow Automation
Demonstrates how AI can safely automate operational workflows by combining multi-agent reasoning, deterministic business rules, SLA assignment, and human oversight.
OpsFlow AI is a multi-agent operations routing system that transforms inbound business messages into structured actions, SLA-aware workflows, routing decisions, and automation-ready payloads.
The architecture combines LLM reasoning, deterministic business rules, safety controls, and human-in-the-loop approval to support real-world operational workflows.
- Python
- OpenAI API
- Multi-Agent Systems
- Workflow Automation
- AI Engineering
- Business Process Automation
- Prompt Engineering
- Structured Outputs
- Human-in-the-Loop AI
- Safety & Guardrails
- Operational AI Systems
- Production-Oriented Architecture
Multi-Agent AI Client Operations Router with Safety, SLA, and Activation Layer
AI lead qualification is only part of the story. Real businesses need to process all inbound communication — sales, support, billing, partnerships, and more.
This project demonstrates a production-oriented AI system that turns unstructured inbound messages into structured actions, routing decisions, and human-review-ready replies.
Transform inbound business messages into:
- structured analysis
- prioritized decisions
- routed operational workflows
- automation-ready payloads
While maintaining:
- safety (untrusted input handling)
- deterministic business logic
- human-in-the-loop control
The system is structured as a set of logical agents:
- Safety Agent (prompt injection detection)
- Intent & Classification Agent
- Entity Extraction Agent
- Routing Agent (deterministic)
- Reply Drafting Agent
- QA / Review Agent
LLM outputs are not blindly trusted.
Routing decisions are enforced via rules:
- Sales → CRM
- Support → Ticket
- Technical incident → Urgent escalation
- Billing → Finance
- Low confidence → Manual review
- Suspicious input → Automation blocked
Each message is mapped to an operational SLA:
- High priority → respond within 1 hour
- Medium → 4 business hours
- Low → 2 business days
- Manual review → no automation
Messages are scored (0–100) using deterministic logic:
- priority
- confidence
- category importance
- urgency signals
- tool mentions
This makes decisions explainable and auditable.
All inbound messages are treated as untrusted input.
Includes:
- prompt injection detection
- message sanitization
- structured output validation
- automation blocking for risky inputs
Before any action:
- replies are reviewed for risk
- unsupported claims are flagged
- human approval is required
The system generates automation-ready payloads for:
- CRM (HubSpot, Pipedrive, etc.)
- Support tools (Zendesk, Intercom)
- Slack notifications
- Internal tasks
Example:
{
"tool": "crm",
"object": "deal_or_task",
"company": "CloudNest",
"priority": "high"
}- inbound sales triage
- support ticket routing
- urgent incident escalation
- billing issue handling
- partnership request routing
- customer success workflows
This system helps companies:
- reduce manual message triage
- respond faster to high-priority issues
- standardize operational decisions
- prevent risky automation
- scale client operations with AI
- maintain human oversight
We are experiencing intermittent outages affecting customer logins. The issue is impacting production users.
| Field | Value |
|---|---|
| Category | Technical Incident |
| Priority | High |
| SLA | 1 Hour |
| Route | Incident Escalation |
| Human Review | Required |
The message is escalated immediately, routed to incident management, and flagged for human review before any external communication.
This architecture can be connected to:
- Make / n8n / Zapier
- OpenAI API
- HubSpot / Pipedrive / Airtable
- Zendesk / Intercom
- Slack / Gmail
- Custom APIs
The full system is implemented in a Google Colab notebook.
👉 Run the notebook to see:
- single message processing
- batch processing
- routing views
- simulated activations
This project is not a prompt demo.
It demonstrates:
- AI system design
- workflow orchestration
- business logic integration
- safety-aware automation
- production-ready thinking
See also:
👉 leadflow-ai — AI lead qualification and scoring system
Together:
- leadflow-ai → acquisition layer
- opsflow-ai → operations layer
MIT