← Selected work
ExperimentsExperiment

Complaint Pattern Detector

I built an n8n experiment that looks for repeated complaints from the same location. Rules control the routing, with an optional LLM step to help interpret a complaint.

Overview

I built an n8n experiment that looks for repeated complaints from the same location. Rules control the routing, with an optional LLM step to help interpret a complaint.

Objective

A single low-severity complaint may only need logging. Several similar complaints in a short period may need escalation. I used synthetic histories to test routing for those cases and for submissions with missing information.

What I Built

  • An importable n8n workflow.
  • Synthetic complaint history and current-submission examples.
  • Expected route outputs for repeatable demonstration.
  • A prompt for optional Claude-compatible analysis through OpenRouter.
  • An offline mock path that does not require an external API key.
  • Routes for logging, missing information, maintenance assignment, pattern escalation, and manual review.

Workflow

Synthetic complaint
  -> required-field validation
  -> location and complaint normalization
  -> recent-pattern lookup
  -> optional AI interpretation
  -> validated route or manual fallback

Current Status

The workflow and synthetic examples are available as a training experiment. They cover routing, repeated-pattern lookup, and fallback behavior. I have not deployed a complaint-management product or measured its operational performance.

Limitations

  • The examples are synthetic and do not measure real operational performance.
  • Similarity normalization is intentionally limited.
  • An LLM recommendation still requires validation and human ownership.

Technology Stack

n8n, Python build scripts, CSV fixtures, Claude-compatible prompting through OpenRouter.

Publication Note

Only synthetic complaint data should be published. No real employee, customer, building, or operational incident information is required for the demonstration.

Checks and current status

Not Yet Evaluated

Workflow assessment

Synthetic complaint scenarios are available; operational performance is untested.

The workflow checks repeated patterns and routes complaints with a manual fallback. Evaluating escalation accuracy would require labeled complaint history and precision/recall measurements.

Demo workflow

AvailableImportable automation and synthetic expected routes exist.

Real-world benchmark

Not availableSynthetic cases do not measure operational performance.

Decision ownership

Human requiredLLM recommendations remain advisory.

Limitations: The experiment is not a deployed complaint-management or incident-detection product.