Project overview · verified local implementation
Support automation
with a human checkpoint.
A local reference implementation that classifies sample support records, retrieves FAQ evidence, drafts a grounded response, and routes sensitive cases to review.
The business problem
Support work often combines repetitive questions with cases that need judgment. A useful system must accelerate routine handling without quietly turning uncertain or sensitive requests into automated outbound actions.
This is an independent portfolio reference implementation using sample records, not a client deployment.
How the system operates
Ingest
Accept a structured support record with an external ID, contact details, subject, and body.
Classify
Apply deterministic category and priority rules with explainable outcomes.
Draft
Match a small FAQ collection and retain the source ID and filename.
Review
Approve, reject, or request changes while recording the decision in SQLite.
Verified scope
Traceable local automation,
not silent delegation.
Sensitive account, privacy, refund, financial, and ambiguous cases require review. Unsupported questions abstain instead of inventing an answer. No email, webhook, help-desk integration, LLM inference, or external delivery is implemented.
Source: public repository README, portfolio documentation, screenshots, and tests.
Engineering journal
Follow the
review-first decision.
The project is a foundation for an article about deterministic triage, evidence retention, and human approval boundaries.
Read Behind the Build ↗