01Detection accuracy auditn=100 · hand-labelled
I hand-labelled 100 LinkedIn posts, 50 written by AI and 50 by people, and ran them through the
detector I built. When it said AI, it was right 21 out of 21 times, and it never flagged a real
person. It also missed 28 of the 50 AI posts. I published the full confusion matrix, plus a
threshold sweep showing why I chose not to trade a first false accusation for one more catch.
labelled test setconfusion matrixprecision vs. recallthreshold sweepsecond-language bias control
The full result set and method, including where it fails.
accuracy audit
02AME NetworksCo-founder · launching Nov 2026
An automated marketing platform for independent insurance agents. From an agent's onboarding answers it generates a unique website on their own subdomain within minutes, matching social covers, and monthly posts, and routes every lead into their CRM. I'm the technical lead: system architecture, 13+ automated workflows, AI generation, hosting, DNS, and payments. About $0.10 per agent per month to run.
Claude APIMake.comCloudflare Workers + KVGoHighLevelStripe
How it works, what I built, and results so far.
case study
A demo site for a fictional agent, generated end to end.
demo site
03VerdahShipped · retired Sep 2026
A Chrome extension that read LinkedIn posts as you scrolled and labelled each one AI, Human, or
Mixed, with a confidence level and a reason. Manifest V3 with no framework. A Cloudflare Worker
held the model key so it never reached the browser, Stripe handled billing, and licenses were
tokens, so there was no login system to maintain. A daily monitor checked that it still rendered.
It passed Chrome Web Store review and went through four production release cycles. I retired it in September 2026 and
shut it down cleanly: subscriptions cancelled, listing unpublished, keys revoked.
Manifest V3Vanilla JSCloudflare WorkersCloudflare KVStripePlaywright
Every post scored as you scroll, with a running summary of the feed. Names blurred.
screenshot
AI, Human, and Mixed verdicts with confidence bands, plus neutral labels for ads.
screenshot
04Oye MilaIn development
A Spanish-speaking voice companion for older adults who live alone. It plugs into the wall and runs
on LTE, so there is no app, no Wi-Fi setup, and no camera. It gives medication and appointment
reminders and real conversation, and texts up to five family members a nightly summary or urgent
alerts. I built the device software, the backend and scheduled jobs, the family portal, and the
bilingual site.
PythonRaspberry PiFly.ioSupabaseTwilio SMSStripe
The product site, with recordings of how she sounds.
oyemila.com
Fully bilingual, written for the families it serves.
oyemila.com/es
05AI-personalized expert programsBuilt in 2 weeks
Turns a coach's or consultant's framework into an automated product: each customer goes through the same method and gets a personalized report and 90-day plan, with follow-up and an upgrade path built in. Deployed for two experts.
GoHighLevelWorkflow automationAI reports
Problem, solution, and what I delivered.
case study
06Academic operations systemPersonal automation
An AI agent workflow that audited four courses through the LMS API, rebuilt grades the platform had wrong, triaged outstanding work against each syllabus, and now sends one daily status text. Recovered 150+ points in about 2 days.
Claude (AI agent)Canvas REST APIAppleScriptiCal
Audit, triage, troubleshooting, and the daily brief.
case study
07Local business digital presenceLive
Took a one-person cleaning business from no website and duplicate stale profiles to a live site, branded email, and a Google Business Profile, without breaking anything that depended on the old domain.
Cloudflare WorkersDNSEmail RoutingLocal SEO
Audit, migration, build, and local SEO.
case study
08Extension policy auditShown on request
A tool that scanned 356 Chrome extensions, read each one's actual manifest file rather than its
store description, and checked it against Chrome's user-data policy. It flagged 12 extensions with
broken or missing privacy policies. The results name real companies, so they stay private. I'm
glad to walk through how it works on a call.
Playwrightmanifest parsingpolicy compliance
09Three practice appsPlayable now
Three browser apps, one HTML file each. No backend, no build step, no account, and your progress
stays in your own browser. Readiness drills the Security+ exam and tells you whether you'd
actually pass. R-Dojo teaches chart reading by scoring how you decided, not whether you won.
Thread runs you through the sales conversations people freeze on.
single-file HTMLadaptive scoringseeded RNG
Readiness. Security+ drill, 186 questions and 8 performance-based items.
open Readiness
R-Dojo. Chart-reading trainer, scored on how you decided.
open R-Dojo
Thread. Sales conversation drills, 16 scenarios.
open Thread