Computer Vision Internship

Detection, classification and the messy reality of real images.

Vision models are easy to demo on clean benchmark images and hard to make work on photographs from the real world. This programme spends its time on that gap: lighting, occlusion, class imbalance and inference cost.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Train and evaluate an image classifier honestly
Run object detection on real footage
Diagnose why a model fails on real-world images
Measure and reduce inference latency

Who it suits

Developers with some Python and ML background who want hands-on vision experience.

What you'll work with

Python
PyTorch
OpenCV
YOLO
torchvision

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Image Classifier with Transfer Learning

~12h estimated

Classify images in a domain of your choosing using a pretrained backbone.

Fine-tune a pretrained network rather than training from scratch. Compare frozen features against full fine-tuning and report the difference in both accuracy and training time.

Assessed against

Frozen vs fine-tuned compared
Augmentation strategy explained
Per-class accuracy, since the average hides failures
A confusion matrix with the top confusions discussed

Suggested datasets

Any public image set — plant disease, waste sorting, medical imaging

You'll finish with

GitHub repo
Training report

Object Detection on Real Footage

~12h estimated

Detect and count objects in video that you did not curate.

Run a detector over real footage — a street, a shop floor, your own recording. Benchmark accuracy will drop, and explaining exactly why is the point of the task.

Assessed against

Real footage, not a benchmark clip
Failure cases identified and categorised
Frames per second measured
A written comparison of benchmark vs real-world performance

You'll finish with

GitHub repo
Annotated output video

OCR Document Extraction

~12h estimated

Pull structured fields out of photographed documents.

Combine detection with OCR and post-processing. Real photographs are skewed, shadowed and creased — preprocessing usually matters more than the OCR engine.

Assessed against

Handles rotation and poor lighting
Field-level accuracy measured on 30+ documents
A confidence threshold below which output is flagged for review
Fails loudly rather than returning a wrong value silently

You'll finish with

GitHub repo
Accuracy breakdown by field

Edge-Ready Model Optimisation

~10h estimated

Make a vision model small and fast enough to run on modest hardware.

Apply quantisation and/or pruning, then measure. Report the accuracy you gave up alongside the speed you gained — a speedup with an unstated accuracy cost is not a result.

Assessed against

Baseline size, latency and accuracy measured first
At least two optimisation techniques applied
The accuracy/speed trade-off stated explicitly
Tested on CPU, not only GPU

You'll finish with

GitHub repo
Before/after benchmark table

Free to apply and free to complete. The completion certificate is optional.

Other domains

AI Agents & Automation

Build autonomous agents that use tools, not chatbots that answer questions.

Generative AI & LLMs

Go past prompt tinkering into retrieval, evaluation and cost control.

Machine Learning

Classical ML done properly — including the evaluation everyone skips.

Ready to Transform Your Enterprise with AI?

Book a free 30-minute consultation with our AI architects. Discover how agentic AI can automate your workflows and multiply your team's output.

No commitment required · Free architecture review · NDA protected

Computer Vision Internship

Detection, classification and the messy reality of real images.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Image Classifier with Transfer Learning

~12h estimated

Classify images in a domain of your choosing using a pretrained backbone.

Fine-tune a pretrained network rather than training from scratch. Compare frozen features against full fine-tuning and report the difference in both accuracy and training time.

Assessed against

Frozen vs fine-tuned compared
Augmentation strategy explained
Per-class accuracy, since the average hides failures
A confusion matrix with the top confusions discussed

Suggested datasets

Any public image set — plant disease, waste sorting, medical imaging

You'll finish with

GitHub repo
Training report

Object Detection on Real Footage

~12h estimated

Detect and count objects in video that you did not curate.

Run a detector over real footage — a street, a shop floor, your own recording. Benchmark accuracy will drop, and explaining exactly why is the point of the task.

Assessed against

Real footage, not a benchmark clip
Failure cases identified and categorised
Frames per second measured
A written comparison of benchmark vs real-world performance

You'll finish with

GitHub repo
Annotated output video

OCR Document Extraction

~12h estimated

Pull structured fields out of photographed documents.

Combine detection with OCR and post-processing. Real photographs are skewed, shadowed and creased — preprocessing usually matters more than the OCR engine.

Assessed against

Handles rotation and poor lighting
Field-level accuracy measured on 30+ documents
A confidence threshold below which output is flagged for review
Fails loudly rather than returning a wrong value silently

You'll finish with

GitHub repo
Accuracy breakdown by field

Edge-Ready Model Optimisation

~10h estimated

Make a vision model small and fast enough to run on modest hardware.

Apply quantisation and/or pruning, then measure. Report the accuracy you gave up alongside the speed you gained — a speedup with an unstated accuracy cost is not a result.

Assessed against

Baseline size, latency and accuracy measured first
At least two optimisation techniques applied
The accuracy/speed trade-off stated explicitly
Tested on CPU, not only GPU

You'll finish with

GitHub repo
Before/after benchmark table

Free to apply and free to complete. The completion certificate is optional.