MakingCode — John DeLeon
The real record of turning a long enterprise career toward AI and automation.
This is what I've actually learned applying 14 years in enterprise software to figure out what comes next — the wins, the wrong turns, and what I still don't know.
Notes from the journey, written as it happens.
Securing a Contact Pipeline: Astro, Cloudflare Pages Functions, Turnstile, and n8n
How I replaced hardcoded mailto links with an edge gateway that verifies humans, signs its own JWT, and hands off to a self-hosted n8n workflow — and the two local-dev gotchas that came with it.
Architecting Idempotent Webhook Pipelines
Why "at-least-once" delivery breaks naive CRM integrations, and the idempotency-key pattern that fixes it for good.
See what I've built.
Bridging the Black Box: Explainable Anomaly Detection for Critical Infrastructure
A DOD/ARL-funded capstone: a Kafka/Spark/Neo4j/TimescaleDB pipeline feeding a two-stage heterogeneous graph transformer that hits 87.8% accuracy detecting attacks on a hardware-in-the-loop water testbed — trained on 369 labeled graph snapshots, a number that should temper every other number in this post.
A five-person, DOD/ARL-funded capstone that got a two-stage heterogeneous GNN to 87.8% accuracy on a physical water-distribution testbed — on 369 labeled graph snapshots, which is the number that should temper every other number in this post.
A Job-Posting Fraud Classifier — and the Gap Between the Model and the Demo
A class-project XGBoost classifier hits 98% accuracy and a 0.985 ROC-AUC detecting fake job postings offline. The deployed demo does not call it — the API route is a coin flip with a TODO comment above it. Here is both halves of that story.
A 98%-accuracy offline classifier, and a deployed demo that never learned to call it — which turned out to be the more useful lesson.
Building a Pickup-Only Order Intake and Mini-CRM for a Cottage Bakery
A tactical breakdown of the pickup-only order intake site and single-admin mini-CRM I built for my wife's cottage-food bakery — and the architectural choices behind it.
Replaced DM/text/notebook order tracking for my wife's bakery with a spam-guarded, server-validated intake flow and a single-admin dashboard she checks once a day.
Hit the same wall translating experience into AI/automation work?
Reply on any post and tell me what actually worked for you. If you're looking to hire instead, see what I work on.