Skills

Run Your Own ML Model On-Device

5 weeks · 0 milestones

Collect your own labelled sensor data, train a small model, and run it live on a microcontroller or single-board computer with measured latency.

Part of: Engineering · Proves for: Engineers, UG

Milestone map

3 milestones

Collect your own labelled data

1–2 weeks

Record your own dataset from a real sensor — sounds, gestures or images — with at least 3 labels. On-device models live or die by their data, so you collect it yourself.

Proof required

Link the dataset (or a summary with sample counts per label) and a short note on how and where it was recorded.

What gets checked

At least 3 labels and 50+ samples each
Recorded by you on the target sensor
Counts per label reported honestly

Common mistakes

Downloading a ready dataset
One label with far more samples than others
All samples recorded in one setting

Powstik Guide

A small dataset you recorded yourself.

Steps

Pick a task with 3+ labels.
Record on the sensor the model will run on.
Vary where and how you record.
Count samples per label.
Hold back a test set before training.

Template

Dataset
Label | samples | where recorded

Test set held back:

What gets sent back

A downloaded dataset.
Fewer than 3 labels.
No held-back test set.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check sample counts per label
Ask how the data varies (people, places, noise)

You'll sign in first, then come straight back here.

Train and test the model

1–2 weeks

Train a small model and test it on the held-back set. Report accuracy per label with a confusion matrix, plus model size and memory, so it can fit on the device.

Proof required

Submit the confusion matrix on your held-back test set, model size in KB, and the training notebook or project link.

What gets checked

Tested only on held-back data
Per-label errors shown, not just one accuracy number
Size and memory fit the target device

Common mistakes

Testing on training data
Reporting one headline accuracy
A model too big for the device

Powstik Guide

A trained model with honest test results.

Steps

Train a small model.
Run it on the held-back set only.
Build the confusion matrix.
Check size against device memory.
Note the weakest label.

Template

Results
Accuracy (test):
Weakest label:
Model size (KB):
Device RAM/flash:

What gets sent back

Results on training data.
No confusion matrix.
No size check.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Confirm the test set was not trained on
Ask which label confuses the model most and why

You'll sign in first, then come straight back here.

Run it live on the device

1–2 weeks

Deploy the model to the microcontroller or single-board computer and run it on live input. Measure how long one prediction takes. Explain it to a reviewer who tests it live.

Proof required

Link a video of live predictions on the device (reviewer supplies at least one input), latency per prediction, and the review log.

What gets checked

Runs on the device, not a laptop
Latency measured
Works on an input the reviewer chose

Common mistakes

Inference actually runs on a PC
Only pre-recorded inputs
No latency number

Powstik Guide

A model predicting live on the device.

Steps

Export for the device.
Flash and run on live input.
Time 20 predictions and average them.
Let the reviewer try their own input.
Log what failed and why.

Template

Review session log
Date:            Duration:
Reviewer name:
Reviewer role and experience:

Challenge 1 (on-device model):
  What they asked:
  My answer:
Challenge 2:
  What they asked:
  My answer:

Change I made after the session:

What gets sent back

Laptop inference.
Pre-recorded demo only.
No reviewer input.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Give it one input of your own
Ask what they would change to cut latency

You'll sign in first, then come straight back here.

Part of

Robotics & Embedded AI →Robotics 4: Perception →

Related outcomes

Milestone map

3 milestones

Collect your own labelled data

1–2 weeks

Record your own dataset from a real sensor — sounds, gestures or images — with at least 3 labels. On-device models live or die by their data, so you collect it yourself.

Proof required

Link the dataset (or a summary with sample counts per label) and a short note on how and where it was recorded.

What gets checked

At least 3 labels and 50+ samples each
Recorded by you on the target sensor
Counts per label reported honestly

Common mistakes

Downloading a ready dataset
One label with far more samples than others
All samples recorded in one setting

Powstik Guide

A small dataset you recorded yourself.

Steps

Pick a task with 3+ labels.
Record on the sensor the model will run on.
Vary where and how you record.
Count samples per label.
Hold back a test set before training.

Template

Dataset
Label | samples | where recorded

Test set held back:

What gets sent back

A downloaded dataset.
Fewer than 3 labels.
No held-back test set.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Check sample counts per label
Ask how the data varies (people, places, noise)

You'll sign in first, then come straight back here.

Train and test the model

1–2 weeks

Train a small model and test it on the held-back set. Report accuracy per label with a confusion matrix, plus model size and memory, so it can fit on the device.

Proof required

Submit the confusion matrix on your held-back test set, model size in KB, and the training notebook or project link.

What gets checked

Tested only on held-back data
Per-label errors shown, not just one accuracy number
Size and memory fit the target device

Common mistakes

Testing on training data
Reporting one headline accuracy
A model too big for the device

Powstik Guide

A trained model with honest test results.

Steps

Train a small model.
Run it on the held-back set only.
Build the confusion matrix.
Check size against device memory.
Note the weakest label.

Template

Results
Accuracy (test):
Weakest label:
Model size (KB):
Device RAM/flash:

What gets sent back

Results on training data.
No confusion matrix.
No size check.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Confirm the test set was not trained on
Ask which label confuses the model most and why

You'll sign in first, then come straight back here.

Run it live on the device

1–2 weeks

Deploy the model to the microcontroller or single-board computer and run it on live input. Measure how long one prediction takes. Explain it to a reviewer who tests it live.

Proof required

Link a video of live predictions on the device (reviewer supplies at least one input), latency per prediction, and the review log.

What gets checked

Runs on the device, not a laptop
Latency measured
Works on an input the reviewer chose

Common mistakes

Inference actually runs on a PC
Only pre-recorded inputs
No latency number

Powstik Guide

A model predicting live on the device.

Steps

Export for the device.
Flash and run on live input.
Time 20 predictions and average them.
Let the reviewer try their own input.
Log what failed and why.

Template

Review session log
Date:            Duration:
Reviewer name:
Reviewer role and experience:

Challenge 1 (on-device model):
  What they asked:
  My answer:
Challenge 2:
  What they asked:
  My answer:

Change I made after the session:

What gets sent back

Laptop inference.
Pre-recorded demo only.
No reviewer input.

Resources

Foundationstart here

Depthgo deeper

What a verifier looks for

Checked by the verifiers you invite to your panel (up to 3). Pick people who can judge the work, not friends.
Give it one input of your own
Ask what they would change to cut latency

You'll sign in first, then come straight back here.