Today we tour real applications, from heart monitors to forests. By the end you will see the same three ingredients in every one of them.
Open with the two best exit-ticket arguments from Class 2, one for moving an app's AI on-device and one for keeping it in the cloud.
Three ingredients. Every time.
Draw these three boxes on the board and leave them up all class. Every application today gets written into them.
Fill the three boxes for these.
sensordecisionreason
smartwatch
light sensor on the wrist (PPG)
→
is the heart rhythm irregular?
→
L · R · P
earbuds
microphone in the earbud
→
was the wake word spoken?
→
E · P
farmer's phone
phone camera, in a field
→
which disease is on this leaf?
→
R
traffic camera
camera above a road
→
how many vehicles passed?
→
B · P
Read the product name, ask the room for the three boxes, then reveal. Accept any reason they can defend.
Start from the sensor.
One spoke per click. For each, ask for one more example from the room before revealing the box.
Each sensor speaks a different shape of data.
Microphone
about 16,000 samples a second
Camera
5 to 30 frames a second
Motion (IMU)
50 to 100 samples a second × 3 axes
Vibration
thousands of samples a second
Environment
one reading every few seconds
ECG / biosignals
a few hundred samples a second
Typical rates; they vary by device. Fast streams like audio and vibration are why Class 8 (signal processing) matters.
Point out the range: from one number every few seconds to sixteen thousand a second. The model, the memory and the battery all follow from this.
The decision box: five kinds of question a model answers.
Keyword spotting, the wake word, is classification on audio. Most TinyML products are classification or anomaly detection: they are the cheapest questions to answer.
Ask which of the five the BrickML uses (anomaly detection) and which the doorbell uses (detection for the person, classification for the keyword).
Now start from the industry.
healthcare
Irregular heart rhythm alert
light sensor · classification
BLERP
agriculture
Crop disease from a leaf photo
phone camera · classification
BLERP
manufacturing
Motor fault, weeks early
vibration · anomaly detection
BLERP
automotive
Driver drowsiness alert
cabin camera · classification
BLERP
smart cities
Vehicle counts, not video
road camera · detection
BLERP
consumer
Wake word in earbuds
microphone · keyword spotting
BLERP
conservation
Chainsaw heard in a forest
microphone · audio classification
BLERP
retail
Empty shelf spotted
shelf camera · detection
BLERP
Highlighted letters are the reasons that keep the AI on the device. Notice how few rows need more than two.
Let the slide sit for a minute while they read. Then pick two tiles and ask the room to defend the letters.
case study · healthcare
A cheap model screens. A careful step confirms.
This shape, a small always-on model waking up a bigger or more careful step, is the cascade pattern. It returns in Class 5.
Be clear that a watch does not diagnose; it flags and suggests. Ask why the screening model must be on the wrist: L (instant), R (no phone nearby), P (heart data is personal), and battery.
case study · agriculture
In the field, offline inference is the whole product.
A farmer photographs a sick leaf. Out here there is often no mobile data.
A model on the phone names the likely disease and suggests what to do next, before the farmer has walked back to the road.
Real systems of this kind exist; the result card here is an illustration, not output from a specific app.
Ask what goes wrong if this needed the cloud: the farmer waits, the photo never uploads, the advice comes too late. Also ask: what happens when the model sees a disease it was never trained on? That question returns in Unit III.
case study · manufacturing
The value is the weeks of warning.
Illustrative curve. The device sends one alert, not twelve weeks of vibration: B, R and E all at once.
This is the BrickML story told as a chart. Ask: who decides where the threshold goes, and what does a false alarm cost a factory compared with a missed one?
team activity · teams of four
Design one of these.
20:00
01
Elephant intrusion alert
warn a village before elephants reach the fields
02
Posture-correcting chair
nudge someone who slouches for too long
03
Earthquake early warning
detect the first tremor and warn seconds ahead
04
Smart traffic signal
change the light by the queue it sees
05
Beehive health monitor
hear when a colony is stressed or queenless
06
Shopping-cart recogniser
identify items as they go into the cart
07
Cough detector in a clinic
count coughs in a waiting room
08
Railway-track crack detector
find cracks from a passing train
Assign cards by team number modulo 8. Twenty minutes: fifteen to fill the worksheet on the next slide, then two teams present in ninety seconds each. Click start on the timer; the arrow keys still work.
Fill five boxes. Nothing else.
sensor
what does it measure?
ML task
which of the five?
output
what happens next?
BLERP, ranked
why on the device?
what could go wrong
one honest failure
The last box matters most. A team that cannot name one way its product fails has not thought about it yet.
Leave this slide up while teams work. Walk the room and push on box five.
one team's answer, to compare with yours
Elephant intrusion alert.
sensor
camera at the village edge, infrared at night
ML task
detection: elephant, or something else?
output
siren, and a text to forest watchers
BLERP, ranked
R · L · B · no network at the forest edge; minutes matter; no video upload
what could go wrong
cattle trigger false alarms; elephants learn to ignore the siren
Show this only after the presentations, or teams will copy its shape. Ask the room which box they would argue with.
our running example
The doorbell, in the same five boxes.
sensor
camera and microphone at the door
ML task
detection (person) and keyword spotting
output
chime; a short clip to the owner
BLERP, ranked
P · B · L · video of a home; no constant upload; instant chime
what could go wrong
a cat counts as a person; a TV says the keyword
Two classes of doorbell now: Class 2 split it between device and cloud; today it gets its sensors and tasks. Next class we open the box and pick the hardware.
exit ticket · on paper, before you leave
Invent one edge AI product for your hometown. Write its sensor, its decision, and the letter that keeps it on the device.
Next class: we open the box. The chips, memory and software that every one of today's products is made of.
The best three hometown ideas make good use-case cards for next year.