Unlearning in Practice

I’m a strong believer in government and civil service workers. Their work makes possible the physical and digital infrastructure on a lot of software relies.

So I was excited to start a partnership with the Michigan Department of Human Services to take the load off of their case workers.

Context: Michigan Pays Caretakers for their Labor

The State of Michigan recently implemented a program to pay people who care for children outside of their immeadiate family.

These caretakers would be paid for work that is often underrecognized since it happens in private homes. Families themselves foot the bill for finding someone to take care of their children. It makes for a place that actively discourages community relationships.

In an effort to fix this, the State allows caretakers to register as official caretakers who would be compensated for their work. It removes some of the financial strain from parents while ensuring a baseline of care and safety for the child.

But Where is the Infrastructure?

Social workers perform multiple visits a day over a range of counties. They document the living conditions of the caretaker to make sure the child is in a safe environment.

This pen-and-ink paperwork is then entered by the social worker by hand into a web interface of the Human Services department. Each visit requires both a manual entry as well as a scanned upload of the physical paperwork. When case workers do multiple visits a day, this work often piles up to quite literally duplicate their work.

Hitting the Road with a Case Worker

We rode along with a few case workers to see what a typical visit was like. I waited in the car for caretaker interviews to respect their privacy, but otherwise sat with the case worker as she filled out 8 pages of paperwork describing the state of the home and person providing care.

This not only ensured that hazardous materials are out of a child’s reach, but even asks about the caretaker’s motivation for providing care. Why do they want to be a caretaker? How long have they been caring for children?

Watching this process (and helping with a few forms myself) taught me two things:

• The assessments provided by these social workers is essential. They have the intelligence and experience to identify a strong provider from a potential risk. This was no place to deploy an ML model where a human’s eye was required.
• Paperwork is no one’s favorite task. And filling in the same question twice is even worse. There had to be a better way to free up case worker time so they can focus on the families.

Our Solution — Woodward: a Digital Assistant for Visit Paperwork

Though we couldn’t save their time on the first two steps of interviewing the provider and writing their answers, we could take the duplicate data entry off of their hands.

When designing a solution, we wanted something that would process the digital tasks in the background, while the provider was actually doing visits. We also knew the solution would have to be accessible to case workers on the road who don’t have ready access to a computer.

The solution looked similar to one I had previously built for a finance team at Amazon. The user would take their handwritten notes and send them via email to an automated inbox.

High-Level Intuition: a Serverless Architecture to Provide Faster Responses

Everytime a user submitted an email to the inbox, the tool would invoke a Lambda function that would automatically process the data. The paper would be scanned for handwritten information which was passed to a NoSQL store for batch processing. The user would receive an email once the first step was completed to let them know their paperwork was received.

The key architecture choice here was deciding what needed to move fast versus what operations could be done a bit slower. Storing the data and responding to the user were the only user-facing tasks that needed to be done immeadiately. Everything after (filling out the digital form, sending it via API) could be done in batches.

The choice of real-time versus batch operations was intentional. Batch operations typically cost less than those required to run immeadiately. It costs would cost less to the client in the long run to separate the batch from real-time operations.

So

Providers

Why it matters: When users ask to be forgotten, models must comply. These notes outline practical steps to implement machine unlearning in deployed systems.

Highlights

• Identifying deletable data and model state
• Efficient unlearning strategies without full retrain
• Observability and audit trails

Takeaway

Unlearning is a product requirement, not a research curiosity. Build it into the lifecycle from day one.