Pipelines, AI agents, and developer tooling I built to keep enterprise documentation accurate without adding headcount. Each starts from a real problem on the MHK documentation team.
After a headcount reduction, one writer was responsible for the full team's output across a 1,900+ article knowledge base.
AI agents that review, score, and publish articles against a custom quality rubric, with structured output schemas, adversarial verification, and the Document360 REST API.
779 new articles in 2026; annual edits up from 88 to nearly 1,000.
Biweekly release notes took a skilled writer a full workday every cycle.
An end-to-end pipeline: Jira JQL extraction, schema enforcement, HTML generation grouped by module, and publishing.
1 day down to 2 hours, across 10+ consecutive on-time releases.
Support handled avoidable contacts, but no one knew which missing or unclear articles caused them.
AI classification and topic clustering over 17,500 support tickets, traced back to specific documentation gaps.
An estimated 875+ support hours saved per year.
Recurring documentation checks were manual and depended on someone remembering to run them.
61 Claude Code agent skills and 3 MCP servers for coverage audits, release note scoring, credential leak scans, and live system variable linting.
Checks that used to be skipped now run on demand, the same way every time.
Documentation lived in Jira, Confluence, and Document360 at once, and the copies drifted apart.
Asynchronous Python pipelines that reconcile and synchronize content across all three platforms with automated schema transforms.
One reconciled set of documentation across three systems.
Thousands of screenshots were locked inside legacy training PDFs and couldn't be searched or reused.
An extraction pipeline that pulled every screenshot and tagged it by UI element, field name, and solution area.
7,541 screenshots made searchable through 34,414 tags.
Sanitized and documented. Each README says what the tool does, what it doesn't claim, and why it exists.
21 Claude Code skills for auditing a 1,900-article documentation corpus and reviewing drafts before they reach customers. Most exist because something went wrong once.
Playwright screenshot capture for documentation, with redaction that fails closed so real patient or customer data never ships in an image.
Sorts documentation into thirteen types from HTML structure alone and explains each decision. No model, no network, microseconds per page.
Drives the real interactive Claude Code CLI from Node through a pseudo-terminal and a headless terminal emulator.
Times a folder of screenshots to a recorded narration by detecting pauses, for fast training videos. ffmpeg only.
Measures real weekly meeting hours from an Outlook calendar and prints them in a timesheet-ready shape. Includes tests and a privacy section.