From messy internal workflows to systems people rely on.

An AI agent that drafts test cases and creates test runs, used daily by a QA team.

PROBLEM

Every ticket reaching QA needed test cases written by hand, almost entirely from information already sitting in the ticket. Three engineers were each losing hours a month to transcription.

APPROACH

An agent that picks up tickets as they enter QA status through the Jira MCP server, reads the full context, drafts test cases for a human to edit or approve, then creates the test run in TestRail through its REST API. Built inside Microsoft Copilot because that is the approved AI runtime, and a tool that needs a security exception does not get adopted.

OUTCOME

In daily production use by three QA engineers. Roughly six hours saved per person per month, now spent on exploratory testing instead of retyping.

• Claude Code
• Microsoft Copilot
• AWS Bedrock
• Jira MCP
• TestRail API
• Human-in-the-loop

Support visibility for bookings that were cancelled and replaced.

PROBLEM

When an order is rebooked the original booking is cancelled and a new one created, but the customer still holds the original confirmation email. Agents had no visibility into the replacement and had to log into the provider platform mid-call to find it.

APPROACH

Interviewed the operations leaders running the workflow and found the cost was the lookup, not the missing data. Prototyped in days, iterated with them, then wrote the specification engineering built from. Search resolves superseded confirmation numbers, so the number the customer reads out still finds the order.

OUTCOME

Live in production. The provider platform lookup is gone from the call path. I validated it in QA and still maintain it.

• Discovery
• Lovable
• Specification
• Production QA

An AI-native platform that turns raw product signals into decisions, specs, and engineering tasks.

PROBLEM

Product teams collect feedback constantly and almost none of it turns into a decision. The signal exists; the path from signal to shipped work is manual and lossy.

APPROACH

A multi-agent pipeline where separate agents own separate stages, coordinated through a durable outbox rather than direct calls. Nothing an agent produces is saved automatically: it proposes, the proposal is validated against a schema, and only then is it written.

OUTCOME

Deployed and running, with an evaluation harness gating every release. Versioned prompt packs, golden datasets, rubric scoring, and calibrated LLM-as-judge graders. Nothing ships without clearing a locked baseline.

• Next.js
• TypeScript
• PostgreSQL
• Anthropic / OpenAI / Gemini
• Playwright
• Evals