Elier Fajardo
All projects

Production for an optical lab

An AI workflow turns photos of handwritten prescriptions into validated orders, and every order is tracked station by station.

Screens
The lab's production board with orders by station
The production board in use. Client and patient names hidden.

Results

  • 6,000+orders processed
  • 210+optical shops served
  • 98.8%field accuracy reading prescriptions
  • <1¢per prescription read

From photo to order

This is how prescriptions arrive: phone photos, handwritten, every shop with its own form. The AI workflow extracts the fields and the team reviews them before production.

  • Handwritten prescription for progressive lenses
  • Work order with sphere and cylinder
  • An optical shop's work order with photochromic material
  • An optical shop's prescription pad with anti-reflective coating
Data extracted by the AI: sphere, cylinder, axis, lens type, material and coating
Production columns: cut and mount, quality control, billing and dispatch

The problem

The lab gets hundreds of orders a week from optical shops across the country. Almost all arrive on WhatsApp as a photo of a handwritten prescription.

Every prescription had to be typed in, and every order was tracked by hand between stations. Knowing where an order was meant asking around.

What I built

A production board where each order moves through the lab's stations and every move is signed by whoever made it.

A service that reads the prescription photo and returns validated fields: sphere, cylinder, axis, add. The team reviews and confirms before production.

How

Strict schemas and human review

The model returns a fixed format that's validated against the lab's manufacturing rules. If something is off, the order doesn't move on its own.

Costs under control

Image preprocessing, provider prompt caching and retries. Each read costs under one US cent.

A reference set run on every change

Real prescriptions, labeled by hand. Every prompt or model change is measured against them before it ships: 98.8% field accuracy.

Got a similar process?

Tell me about it. The first call is free.