Nelson Ireson

Projects

Things I have built for the newsroom

More about how I build
Planning board assignments Workflows one platform dedup, retries KBB site Autotrader site Team alerts publish status errors

Editorial-to-CMS publishing platform

One platform that replaced a tangle of third-party automations between the editorial planning board and the KBB and Autotrader CMSs. Two-way status sync, revisions, video assets, retries with backoff, and errors an editor can read.

MVP delivered solo in two weeks, then built out through the spring and summer. In daily production use since November 2026.

Sources OSINT, filings collect de-duplicate already covered? sort by assignability checked against the published archive Chat digest twice a day Briefing detailed findings

News-finding tool

Collects stories from across the industry, researches them against open-source intelligence and primary documents, removes duplicates, checks each one against what the newsroom has already published, and sorts what is left into categories an editor can assign from. Twice a day it posts a digest to the newsroom’s chat, with a briefing document for the detailed version, and it tracks what has been assigned and what has been completed.

Yields 30 or more news assignments a week.

Every live page, scanned on a schedule flags Report wrong-year photo missing pricing broken link

Content-integrity scanner

Scheduled defect detection across tens of thousands of live pages: wrong-year photos, empty placeholders, missing pricing, leaking CSS, broken links. One detector was retired after a hand-labeled sample found nothing real.

Hundreds of offline tests, which also caught flaws in the project’s own acceptance criteria.

AI assistant MCP connector human approves spend shared monthly cap returned text isolated Vendor API paid data query paid call untrusted

Governed MCP connector

Gives the newsroom’s AI assistant access to a paid third-party data API, with human approval for spend, a shared monthly cap, isolation of untrusted returned text, and one-click install with encrypted key storage on macOS and Windows.

More than a thousand tests. Three rounds of blind review. Measured the real budget need before anyone signed a contract, and it came in well under the estimate.

I build by directing AI coding agents against my own designs, requirements, tests and reviews, and I say so plainly. The judgment, the specification and the verification are mine.

How I build

I write the design, the requirements and the acceptance tests first, then direct AI coding agents against them, then review and verify the result myself. The agents type; the judgment, the specification and the verification are mine. Everything above runs in production for a newsroom of about thirty editors, and the test counts are real.

Outside work I run a production automation server, a self-hosted inference stack on Apple Silicon, an AI-assisted video pipeline for drone and automotive footage, a chess-analysis app for Android, and a handful of embedded projects, from a GPS dog-tracker collar to an in-car data logger. Details on request, or on LinkedIn.