Building AI that works in the real world
Big idea: building good AI means testing it where it will really be used, knowing when it’s unsure, and working with the people who know the job.
The story in short
At the school hackathon, Bunny and his friends built Kasa Mitra (“rubbish friend”), an app that sorts waste from a photo. They collected 1,200 photos, with consent, from 20 families, labelled each one twice, and sealed 200 as a test set: 192 of 200 right (96%). They borrowed a model already trained on millions of photos and made it small enough to run on a phone. Then, in the school canteen, it got only 31 of 50 right (62%): steel plates, banana leaves and tube lights looked nothing like home. Savitri akka, a pourakarmika with 19 years of sorting experience, found 11 wrong labels and helped take 600 new photos. Now, below 85% confidence, the app says “ask Savitri akka”, and the canteen score rose to 45 of 50.
The idea
- Labelled dataset and a held-out test set the model never trains on.
- Transfer learning: start from a model trained on a big task, then teach it a new one.
- On-device AI: runs on the phone, with no internet, better for privacy and energy.
- Distribution shift: the real world looks different from the training data.
- Label noise: wrong labels, best caught by a domain expert.
- Confidence threshold + human in the loop: when the AI is unsure, a person decides.
- Privacy by design: protecting people’s data from the very first step.
Bunny’s eight questions
The six from Volume 1 (what did it learn from, who might it be unfair to, how sure is it and who checks, what happens to our data, can we say no, who is in charge), plus: can it be tricked? and is it worth the cost?
Remember: the smartest machine still needs a wise human beside it.
Check yourself
A face-unlock that worked perfectly indoors fails in bright sunlight. What is this?
Why did the team "borrow" a model already trained on millions of photos?
Kasa Mitra says "ask Savitri akka" when it is less than 85% sure. What is this design?
More for you: 🎒 Students · 🧑🏫 Parents and teachers