rAs Kids📄 📱 📊 ⚠️ 📸
🎒 What you need
- the shared phone, your trained model, and your Python code
- one real person willing to describe a problem they actually have
- paper or card for a one-page report
✋ Do this
- Interview one real person — a farmer, a trader, a shopkeeper, a nurse — and write down one classification they make by eye every day. Their words, not yours.
- Collect your own dataset for it, deliberately including awkward conditions: shade, dust, different backgrounds. Aim for 60 images across your classes plus a 'none of these' class.
- Train the model, then split off ten items it has never seen and build a confusion matrix from them. Report the real numbers, including the bad ones.
- Run your module 5 attacks and write a LIMITATIONS section: at least four specific conditions under which this model should not be trusted.
- State a threshold rule in one sentence, such as: 'Below 80% confidence, this tool must not be used to decide anything — ask a person.'
- Lay out your one-page report: the problem in the person's own words, how it works, the matrix, the limitations, the threshold. Photograph it. Then show it back to the person you interviewed and write down what they say.
💡 Why it works
The limitations section is the part that separates a school project from professional work. Anyone can post an accuracy number; stating precisely where your system must not be trusted is what makes it safe to use, and it is exactly what a model card in industry contains. You have now built, tested, attacked, bounded and documented a machine learning system end to end.
🔥 Challenge
Give your report to someone who has never met you and ask them: 'Would you rely on this, and why not?' Their objection is your next project.
📖 New words
model carda short document stating what a model does, how well, and where it fails
deploymentputting a model to work on real decisions with real consequences
limitationa stated condition under which your system should not be trusted
🎖️
Honest Engineer
Tap when you have finished this module.