Concept build
Private-AI Support Inbox Router
A self-hosted Llama 3.2 model triages every customer email: it answers FAQs in the same thread and escalates refunds, damage claims and angry customers to a person, without sending email content to a third-party AI.
Verified in the code
- intent categories
- 5intent categories
- escalation guardrail keywords
- 24escalation guardrail keywords
- emails sent to cloud AI
- 0emails sent to cloud AI
- workflow nodes
- 12workflow nodes
The problem
Support inboxes mix easy questions with urgent complaints. Reading everything by hand is slow, and many businesses don’t want customer emails sent to an outside AI service.
What we built
A workflow polls the inbox, classifies each email with a local model into five intents, and acts: in-thread replies for FAQs and order status, a label and an admin hand-off for anything sensitive, and silent marking for spam. Rule-based guardrails override the model for refunds, damage and legal topics.
Capabilities
What it does
- Runs Llama 3.2 locally, so email content never goes to a third-party AI API
- Five-way classification with a strict JSON output contract
- Defensive parsing: falls back safely and escalates to a human when output is unclear
- Keyword guardrails force human review for refunds, damage, missing items and legal topics
- Replies inside the customer’s original thread
- Escalations are labelled and forwarded to an admin with a summary
Under the hood
Routing
- 1New email
- 2Local Llama 3.2 classification
- 3Guardrail override
- 4Reply in thread · or · Label + escalate
Stack
Want something like this for your business?
Book a free call. We’ll look at how your business runs and tell you honestly what’s worth automating, and what isn’t.