Status: In Development

View the source code on GitHub

Problem

Support teams regularly process repetitive requests that still require careful interpretation, policy awareness and human judgment.

A useful AI operations system should help operators analyze requests without silently inventing facts, bypassing review or taking autonomous customer-facing actions.

Current solution

The current milestone accepts clearly synthetic support requests and produces a validated, machine-readable recommendation containing:

  • Category
  • Urgency
  • Rationale
  • Missing information
  • Suggested reply draft
  • Security flags
  • Validation errors
  • Mandatory human-review status

Every result retains:

{
  "human_review_required": true
}

Architecture

Synthetic support request

Input validation

Deterministic analysis

Output validation

Human Review Guard

Reviewable structured result

Two implementations currently share the same documented contract:

  • A dependency-free Python reference implementation
  • An inactive, manual-only n8n workflow using built-in nodes

The Python implementation remains the canonical parity-testing reference for the n8n workflow.

Safety design

The current version is deliberately limited and does not send messages or connect to production systems.

Implemented controls include:

  • Synthetic data only
  • Strict input and output contracts
  • Mandatory human review
  • Prompt-injection detection
  • Human-review bypass detection
  • Protected-field manipulation detection
  • Safe rejection of malformed input
  • No credentials committed to the repository
  • No external APIs or LLM calls
  • No autonomous actions

Verified results

The repository currently includes:

  • 22 passing automated tests
  • Python and n8n workflow parity tests
  • Normal, urgent, incomplete and ambiguous fixtures
  • Prompt-injection regression fixtures
  • Malformed-input and tampering tests
  • Successful native workflow execution
  • Successful export and re-import round trip

Native workflow verification was completed on self-hosted n8n 2.14.2.

Compatibility with newer n8n versions has not yet been verified.

Current milestone

Completed:

  • Task 001 — Local classifier and validation baseline
  • Task 002 — Native n8n workflow integration

The workflow remains inactive and manual-only.

Known limitations

  • Classification is currently deterministic and keyword-based.
  • Urgency detection depends on explicit indicators.
  • Missing-information rules are illustrative.
  • Suggested replies use controlled templates.
  • No knowledge base or retrieval layer is connected.
  • No LLM is currently used.
  • The system has not been tested with real customer data or production traffic.
  • Compatibility is currently verified only on n8n 2.14.2.

Next steps

Planned development includes:

  1. LLM-assisted analysis behind the existing validation contract
  2. Grounded knowledge-base retrieval
  3. Human approval and decision logging
  4. Evaluation metrics and regression reporting
  5. A small demonstration interface

The project will continue to avoid autonomous customer-facing actions and real customer data during development.