Build a Simple Neural Network to Read Handwritten Postal Codes
The situation
The provider (around 1,200 employees, around 900,000 parcels per week) currently pays its vendor about INR 0.40 per parcel for OCR; a viable in-house model would target under INR 0.10 fully loaded.
Your task
Train and evaluate a small neural network for handwritten digit recognition that the operations team can compare to a vendor OCR module.
The brief
You start with
You receive a labeled dataset of about 60,000 handwritten digit images (28x28 grayscale) drawn from Indian postal forms.
What you do
- Build two models from scratch in PyTorch: (1) a 2-layer multi-layer perceptron baseline, (2) a small convolutional neural network (CNN).
- Train both with proper validation splits and regularization (dropout or weight decay).
- Report top-1 accuracy on the holdout set, confusion analysis, and per-digit recall (since a 1-in-10 digit error in a 6-digit code is a full mis-routed parcel).
- Write a 2-page brief comparing your CNN result to a quoted vendor accuracy of 99.2%.
What you will practice
- Implement a feed-forward and a convolutional neural network from scratch
- Use validation splits, early stopping, and regularization correctly
- Reason about per-class errors instead of relying only on top-1 accuracy
- Translate a model evaluation into a business comparison
A suggested way in
- Start with the MLP to get a baseline number before the CNN
- Use torchvision transforms for simple data augmentation
- Plot per-digit recall — digit 7 being confused with 1 is the classic failure mode
Skills you will practice
Neural networks, Convolutional neural networks, Pytorch, Regularization, Model evaluation, Python
Where it leads
Machine Learning Engineer, Computer Vision Engineer, AI Engineer
How it is assessed
- Model accuracy35%
CNN beats MLP and approaches the vendor's quoted 99.2% on the holdout set.
- Methodology30%
Proper train/val/test split, regularization used and justified.
- Per-class analysis20%
Per-digit recall reported and discussed in operational terms (mis-routed parcels).
- Business brief15%
2-page brief compares in-house vs. vendor with cost and risk discussion.
Resources
- PyTorchTool
Standard deep-learning library for educational and production CNN work.
- PythonTool
Glue language for the training script and evaluation.
- scikit-learnTool
Confusion matrices and classification reports.
- MatplotlibTool
Visualize per-digit recall and training curves.
- JupyterTool
Notebook for the train + evaluate loop.
- NotionTool
Write the 2-page business brief.
Finish it. Earn this.
A verified certificate with your name. Anyone can check it.
Time to practice.
Solve this challengeMore like this.
- Software EngineerBuild a Multi-Level Feedback Queue Scheduler in xv6
- Software EngineerReverse-Engineer a Stripped Binary to Restore Linking Symbols
- Software EngineerImplement a Distance-Vector Routing Algorithm
- AI EngineerEnd-to-End Lane Following on a Donkeycar Platform
-
Build a Multi-Level Feedback Queue Scheduler in xv6 Software Engineer Fork the xv6-riscv repository and implement MLFQ in proc.c with 3 priority queues, per-queue time slices (10ms / 20ms / 40ms), priority demotion on sl operating systemsschedulingc programming Open challenge → -
Reverse-Engineer a Stripped Binary to Restore Linking Symbols Software Engineer Receive the stripped shared object (Linux ELF, x86-64, ~280KB), a crash dump with 12 frames of indecipherable addresses, and 4 years of partial git hi reverse engineeringlinking and loadingmachine code Open challenge → -
Implement a Distance-Vector Routing Algorithm Software Engineer Implement distance-vector routing in Python: each simulated router maintains a distance vector, exchanges vectors with neighbors every 1 second of sim routingdistance vectornetwork simulation Open challenge → -
End-to-End Lane Following on a Donkeycar Platform AI Engineer Use the public Donkeycar Tub dataset (or collect about 30 minutes of driving on the simulator). Train a CNN-policy baseline (the Donkeycar default arc end to end learningimitation learningpytorch Open challenge →