ArcVault AI Intake & Triage Pipeline

AI-powered support ticket classification, enrichment, and routing pipeline.
Built for the Valsoft AI Engineer Technical Assessment.

What It Does

Processes unstructured inbound support messages through a 6-step agentic workflow:

1. Ingestion — accepts raw messages (webhook, file, or direct input)
2. Classification — GPT-4o-mini assigns category, priority, and confidence score
3. Enrichment — GPT-4o-mini extracts summary, entities, and urgency signal
4. Routing — pure Python maps classification to destination queue
5. Structured Output — writes output/results.json + POSTs to webhook.site
6. Escalation — flags low-confidence or high-risk tickets for human review

Quick Start

Python Pipeline

# 1. Install dependencies
pip install -r requirements.txt

# 2. Set up environment
cp .env.example .env
# Edit .env — add OPENAI_API_KEY and webhook.site URLs

# 3. Run
python -m apps.pipeline.run

Output: output/results.json with all 5 processed records.

n8n Visual Workflow (Cloud — Live)

The latest workflow is deployed and active on n8n Cloud:

https://youssefc.app.n8n.cloud/webhook/arcvault-intake

Send any ticket directly — no setup required:

curl -X POST https://youssefc.app.n8n.cloud/webhook/arcvault-intake \
  -H "Content-Type: application/json" \
  -d '{"id":1,"source":"Email","raw_message":"Login broken after Tuesday update. 403 error. Account: arcvault.io/user/jsmith"}'

All 5 sample inputs are in demo/send_tickets.txt — paste any line directly into PowerShell:

# Ticket 1 — Bug Report
curl.exe -s -X POST "https://youssefc.app.n8n.cloud/webhook/arcvault-intake" -H "Content-Type: application/json" -d "{\"id\":1,\"source\":\"Email\",\"raw_message\":\"Hi, I tried logging in this morning and keep getting a 403 error. My account is arcvault.io/user/jsmith. This started after your update last Tuesday.\"}"

n8n Visual Workflow (Self-Hosted)

To run the workflow locally instead:

# 1. Start n8n
cd infra && docker compose up -d

# 2. Open http://localhost:5678

# 3. Import workflow
#    Workflows → Import from file → infra/n8n/workflow.json

# 4. Set up credential
#    Settings → Credentials → New → Header Auth
#    Name: "OpenAI Header Auth"
#    Header Name: Authorization
#    Value: Bearer sk-your-key

# 5. Update webhook.site URLs in the Standard Queue and Escalation Queue nodes

# 6. Activate the workflow

# 7. Test
curl -X POST http://localhost:5678/webhook/arcvault-intake \
  -H "Content-Type: application/json" \
  -d '{"id":1,"source":"Email","raw_message":"Login broken after Tuesday update. 403 error. Account: arcvault.io/user/jsmith"}'

Run Tests

python -m pytest tests/ -v

Project Structure

arcvault/
├── apps/pipeline/       # Python pipeline package
│   ├── run.py           # Entry point
│   ├── models.py        # Pydantic schemas
│   ├── prompts.py       # LLM prompts
│   ├── classifier.py    # Classification step
│   ├── enricher.py      # Enrichment step
│   ├── router.py        # Routing + escalation
│   └── outputs.py       # JSON + webhook output
├── infra/
│   ├── docker-compose.yml
│   └── n8n/workflow.json
├── output/results.json  # Generated output
├── tests/               # pytest tests
└── docs/                # Architecture + prompt docs

Model Choice

GPT-4o-mini — chosen for consistent JSON mode output, low cost (<$0.01 for 5 inputs), and strong classification accuracy for short support messages. Temperature set to 0.1 for reproducibility.

Escalation Criteria

A ticket is flagged for human review if any of the following are true:

• Classification confidence < 70%
• Message contains: outage, down for all, multiple users affected
• Billing Issue with amount > $500
• High priority + confidence < 80%

See docs/architecture.md for full rationale.