What OCR Actually Gets Wrong on Handwritten Recipes
OCR demos always use a clean printed card. Real recipe boxes are faded and handwritten, where OCR fails in ways that quietly ruin the dish. Here's what breaks, and what actually fixes it.
Read moreThe Chronicle
Thoughts, tutorials, and tales from the digital frontier. Lessons learned and wisdom shared.
Everything That Broke Along the Way
The commit history tells a story. Not the one I planned. The one where every assumption I made about backups got corrected by production.
Read moreRestores Are Harder Than Backups
A backup that can't be restored is a zip file you feel good about. Building the restore side taught me how many ways a restore can succeed while failing.
Read moreThe Pipeline Strategy Pattern
The first version had a big switch statement. Every strategy duplicated the compression logic. Then I realized what I actually had were composable steps.
Read moreI Built My Own Backup System
I had a Node app that logged successful backups to Discord. Failures were silent. I found an eleven-day gap by accident. So I built something better.
Read moreFrom Inbox to Database Without a Human in the Middle
Weekend experiment is over. This is what happens when you put it to work. Invoices and purchase orders arriving by email, dozens a day, and someone spending half their day typing fields into a system. Email to S3 to Lambda to done.
Read moreWhy I Built Flour Power
Every recipe app I tried was designed for the same thing: browsing. Scrolling. Saving links. They were good at collecting recipes and terrible at the part where you actually cook them. Cooking isn't reading. Cooking is a sequence of timed actions where your hands are busy and your attention is split between three things on the stove.
Read moreLess Than a Penny Per Document
People hear "I replaced my OCR pipeline with a vision model" and the first thing they ask about is cost. Fair question. I assumed it would be expensive too. Under a penny per document.
Read moreOne API Call Changed Everything
Sunday morning. I'm close to just giving up and typing everything by hand. But I want to try one more thing. Instead of OCR to get characters, what if I just ask a vision model what the document says?
Read moreWhen the Cleanup Code Becomes the Project
Tesseract can't handle the handwriting. Time to spend some money. AWS Textract has built-in handwriting detection. It's better. But "better" and "good enough" are different things.
Read moreStay on the Path
Get notified when new articles are published. No spam, just valuable content.
Prefer a feed reader? Subscribe via RSS