Sidecar

Chrome extensions, browser automation, and small internal tools. Built and shipped by one person. Every claim below links to the thing it describes — if a line has no link, it does not go in a message.

Claims on this page —

Shipped

01
live

Verdah

A Chrome extension that scores LinkedIn posts for AI-likelihood and puts a badge on each one. Manifest V3, no framework, Cloudflare Workers proxy holding the model key, Stripe billing, and license-token auth with no login system to maintain.

Manifest V3Cloudflare WorkersCloudflare KVStripeVanilla JS
It is live and it takes payment. verdah.com The permission scope and data handling are written down publicly. verdah.com/privacy.html It passed Chrome Web Store review and is at v1.1.1 after four release cycles. chromewebstore.google.com/detail/pjejfbp…
02
measured

Verdah accuracy audit

I labelled 100 posts by hand, 50 AI and 50 human, and ran them through my own product. It matched on 54. It called 48% of the AI posts human. What it does do well is stay quiet: one false accusation in 50 human posts, and 89% precision when it does say AI. I publish this because most detectors publish a number nobody can check.

labelled setconfusion matrixn=100
The full result set and method, so the numbers above can be checked. sidecar-59t.pages.dev/accuracy.html
03
daily

verdah-monitor

A Playwright job that opens the live store listing in an authenticated browser every morning at six, checks the page actually rendered rather than just returned 200, and alerts on three channels when it hasn't. This is the difference between monitoring and pinging a URL.

PlaywrightcronSMTP + iMessage alerts
Source, including the render check rather than a status-code check. github.com/yazz-prog/verdah-monitor
04
tool

sidecar-scout

Scans the Chrome Web Store, downloads the CRX for each extension, unpacks it, reads the real manifest.json, and scores the declared permissions against Chrome's Limited Use Policy. It reads what an extension actually ships, not what its listing says. 356 extensions scanned in one pass.

PlaywrightCRX unpackingpolicy scoring
Source and the scoring rules it applies. github.com/yazz-prog/sidecar-scout
05
tools

Three training apps

Single-file browser apps, no backend, no build step. Readiness drills CompTIA Security+ with adaptive weighting and blueprint-proportional exam mode. R-Dojo is a price-action trainer that scores process rather than outcome and tracks five skills across sessions. Thread drills client-acquisition conversations across sixteen scenarios.

single-file HTMLlocalStorage stateadaptive scoring
Readiness, playable in a browser. sidecar-59t.pages.dev/apps/readiness.html R-Dojo, playable in a browser. sidecar-59t.pages.dev/apps/rdojo.html Thread, playable in a browser. sidecar-59t.pages.dev/apps/thread.html

What I do for clients

A

Extensions

Build a Manifest V3 extension, fix one that got rejected, or take over one nobody is maintaining. I have taken an extension from empty folder through store review to taking money, so I know where the review process actually bites.

B

Browser automation

Get data out of a web app that has no export button, on a schedule, with a monitor that tells you the morning it breaks instead of the month you notice. Scrapers without that second part are a liability you inherit.

C

Small internal tools

The spreadsheet that three people maintain by hand, turned into something that maintains itself. Usually a few hundred lines and a scheduled job, not a platform.

What I don't claim