I make AI actually work in support operations.

AI QA Scorer: transcript scoring and coaching

Problem

Support teams burn hours listening to calls and hand-scoring transcripts against a rubric, and the scores come out inconsistent.

Stack

React client UI · OpenAI Responses API (gpt-5.6-luna) · strict JSON Schema output · fixed rubric scoring in TypeScript · Cloudflare D1 for rate limits.

Result

Scores a pasted transcript almost instantly, roughly 5-10x faster than manual review (up to ~90% less review time on routine calls), and applies the rubric the same way every time. Based on the production version of this workflow, which cut overall QA-review time about 40%.

ReviewSignal: customer review trend intelligence

Problem

Product-quality and safety patterns stay invisible when teams only look at contact-center data and never mine the review firehose.

Stack

React dashboard · TypeScript data models · client-side state · Tailwind + custom CSS · simulated review data.

Result

The original production process (built at Helen of Troy) chewed through tens of thousands of reviews and surfaced emerging issues no single review shows, including a fan fire-report pattern in Canada and unusual-smell reports in air-purifier reviews.

This public demo is a portfolio-safe rebuild on simulated data.

AI support email triage with human approval

Problem

A small consumer-products support team is buried in repetitive order-status, product, refund, and complaint emails.

Stack

n8n · webhooks · JavaScript code nodes · Switch routing · OpenAI Responses API (gpt-4.1-mini) · structured JSON output · human approval via Wait node · Google Sheets logging · V2: Gmail, Slack, and Jira test integrations with audit logging.

Result

In prototype testing, about 40% of routine inbound emails got a usable first draft. Every draft still requires human approval before sending.

SupportOps: one policy set for chat and MCP

Problem

Most support chatbots answer confidently whether or not they actually know. This project gives people a cited chat UI and gives AI clients the same synthetic policy set through a tool-scoped MCP server.

Stack

React chat UI · TypeScript API routes · OpenAI Responses API · MCP TypeScript SDK with stdio transport · source citations · strict tool schemas · local audit logs · Cloudflare D1 for chat logs and handoff tickets.

Result

Routine policy questions return a cited answer in the web chat or through an MCP client. Uncertain cases stay visible for human review.

The MCP download includes four allowlisted tools. Three are read-only. The one write tool saves a local demo escalation.