04 — NOTES

Things I've had to figure out.

Long-form versions of what I build — dedup strategies, RAG isolation, and the infrastructure underneath four products.

2026-01

Deduplicating an automated news pipeline — title matching isn't enough

Why comparing headlines catches almost nothing, and what to check instead when the same story shows up from five different outlets.

DATA

2026-02

Multi-tenant RAG: keeping one tenant's documents out of another's answers

Row-level tenant IDs in Postgres don't protect you once a vector store enters the picture. Isolation has to happen at retrieval, not just at the database.

RAG

2026-02

Extracting geolocation from image EXIF at scale, and why the metadata lies

Phones and cameras write location metadata inconsistently. Extraction without a validation pass is extraction you can't trust.

DATA

2026-03

A proxy API layer: how to let a frontend use third-party APIs without shipping keys

If your frontend calls a third-party API directly, your API key is in the browser. Here's the pattern that keeps it server-side.

INFRA

2026-03

Running four products on one VPS with Dokploy and Docker — the actual setup and monthly cost

The real infrastructure behind everything I ship: reverse proxy, container layout, secrets, backups, and what it actually costs per month.

INFRA

2026-03

What shipping to a wholesale distributor taught me that no tutorial did

MyLedger went live with a real, non-technical business. The lessons that mattered weren't about the tech stack.

DATA