Deploying Handwriting Recognition at Scale for the State of Michigan

I’ve always believed in the value of government and civil service workers. Their daily efforts make possible the physical and digital infrastructure on which so much of our society—and our software—relies.

That’s why I was excited to partner with the Michigan Department of Human Services to explore how technology could ease one of the most frustrating parts of their job: paperwork.

Context: Paying Caretakers for Their Labor

Michigan recently launched a program to compensate people who provide child care outside of their immediate family.

This matters. Much of this work has long gone unpaid and unrecognized, even though families often struggle to find and fund quality child care. The program relieves some financial strain on parents while creating a safer, more supportive environment for children.

But programs like this require infrastructure. And that’s where we ran into problems.

The Paperwork Bottleneck

Case workers make multiple visits a day across different counties. At each visit, they document everything—from safety checks (Are hazardous materials out of reach?) to personal questions (Why do you want to be a caretaker?).

These assessments are essential, but the process is inefficient. Workers fill out detailed forms on paper, then later re-enter the same information into a state web portal. The result? Hours lost to duplicate data entry.

Hitting the Road with a Case Worker

To understand the problem firsthand, I rode along with several case workers. I watched one fill out eight pages of forms during a single home visit.

That experience taught me two things:

• Human judgment is irreplaceable. These workers bring the expertise to tell a supportive home from a risky one. An ML model should never replace that.
• Paperwork wastes time. Duplicating answers drains energy from the real work—supporting families.

Our goal became clear: use technology to remove the duplicate data entry while keeping the human assessment intact.

Our Solution: Woodward, a Digital Assistant

Enter Woodward—a lightweight system that digitizes handwritten visit forms.

Case workers simply email their handwritten notes to a secure inbox. Woodward scans the forms, extracts text using handwriting recognition, formats the data according to state requirements, and returns a draft for review. With one click, workers confirm or correct the entry before it’s automatically uploaded to the state portal.

The design was inspired by a data-processing tool I had previously built at Amazon. The key was keeping the workflow simple, mobile-friendly, and fast enough to run while case workers were still on the road.

Under the Hood: A Serverless Pipeline

Every submission triggers a Lambda function that scans the document, extracts handwritten fields, and stores them in a NoSQL database for batch processing.

The pipeline works in three stages:

• Intake: Forms are received, scanned, and confirmed as “in process.”
• Recognition: Text is parsed and mapped to the correct state database fields, even accounting for quirks like scribbled-out bubbles.
• Review & Submit: Case workers get an email with the draft. A “yes” pushes the data via REST API to the state; a correction opens the web portal for edits.

Impact: More Visits, Less Strain

We piloted Woodward with three case workers over one week. The results were immediate:

• Faster workflows: Forms submitted within 20 minutes of a visit’s end.
• Less redundancy: Workers reported fewer cramped wrists and more time with providers.
• Higher productivity: Daily visits increased by 12% week-over-week.

Accuracy improved with fine-tuning. At first, Woodward struggled with one worker’s handwriting style, recognizing only 62% of his forms correctly. After additional training on his writing samples, overall recognition rose to 96% across 28 case workers.

The remaining 4% of errors were fed back into the model for further improvement.

Takeaways: Caring for Caretakers

Technology can’t replace the expertise and empathy of social workers—but it can give them more time to focus on families instead of forms.

With Woodward, we saw firsthand how machine learning can serve the public good by reducing administrative burdens for the people who support our communities every day.