Projects

Things I've designed and built: robots, deep learning networks, and embedded systems.

Autonomous cup pong robot

Jan–Apr 2026

Embedded system with LiDAR object detection

The robot's purple 3D-printed frame, launch tube and twin flywheel motors Breadboard wiring for the ESP32, LiDAR and motor drivers
  • Reached 100% target detection and ±2° aim accuracy by reading a TF-Luna LiDAR over 400 kHz I²C during a servo-driven 115° polar scan.
  • Cut integration failures by building 7 isolated PlatformIO test environments in C++ to validate each peripheral (LiDAR, servo, hall sensor, IR beam, motor drivers) before bringing the system together.
  • Hit a 200 cm launch range by generating multi-frequency hardware PWM for the dual flywheel and indexer motors, with a staged duty-cycle spin-up to overcome static load before firing.
  • C++
  • ESP32
  • I²C
  • PlatformIO
  • Motor control
  • Hardware-in-the-loop

Garment classifier

Jan–Apr 2026

Convolutional neural network approach

A first step toward a clothes-folding robot: a model that looks at a single garment and tells which of six types it is. Read the final report (PDF).

  • Reached 92.8% test accuracy and 0.928 macro-F1, against 0.899 for a HOG + SVM baseline, by fine-tuning ResNet-18 with a custom fully connected head (8.5M trainable parameters).
  • Raised the weakest class (skirts) from 0.74 to 0.90 F1 by merging 3 datasets (6K+ images) and correcting an 8:1 class imbalance through capping and augmentation.
Model diagram: input image, ResNet-18 residual stages, then a custom linear head that outputs the predicted class
ResNet-18 backbone with a custom fully connected head.
Grid of 18 clothes from my closet, each labelled with predicted and true class; all correct
Tested on clothes from my own closet: 18 of 18 correct.
Confusion matrix on the test set across hoodie, long pants, long sleeve sweater, shorts, skirts and t-shirt
Confusion matrix on the 900-image test set.
Images per class across the three merged datasets
Images per class across the 3 merged datasets.
Precision, recall and F1 per class before and after balancing
Per-class metrics before (left) and after (right) balancing the classes.
  • PyTorch
  • CUDA
  • Hyperparameter tuning
  • Data preprocessing

Mail delivery robot

Sep–Dec 2025

Colour-detection based, line-following rover

Heatmap of the robot's belief over offices 2 to 12 across 13 time steps, starting uniform and converging to office 8 with about 0.87 probability
The Bayes filter's belief over time: it starts spread evenly across every office, tracks the robot as it moves, and settles on office 8.
  • Reached over 90% localization confidence across an 11-state topological map with only 4 distinguishable landmarks, using a discrete Bayes filter with probabilistic motion and measurement models.
  • Eliminated false landmark detections under changing light with a weighted colour classifier and an 11-frame debounce filter on camera measurements.
  • Completed autonomous deliveries to 3 random targets from an unknown start by splitting PID line-following and localization into separate ROS nodes talking over pub/sub topics.
  • ROS
  • rospy
  • Bayes filter
  • PID
  • Python
  • Linux

DC power supply

Sep–Dec 2025

Fully functional power supply built from scratch

Front of the board: LCD screen, rotary knob and button, held in 3D-printed clips My hand-soldered op-amp board mounted on the back of the main PCB
The supply running on a lab bench, screen showing 4.77 V against a 5.00 V target Side view of the stacked boards
  • Designed a two op-amp amplifier that turns a 0–2.5 mA sensing current into an adjustable 0–19 V output.
  • Built and tested an op-amp filter in SPICE that cuts noise above 50 Hz by 92%.
  • Designed a single-via PCB in Eagle with test pins, then fabricated, soldered and validated it with a multimeter and oscilloscope, landing within 1% of a lab bench supply.
  • PCB design
  • Eagle
  • SPICE
  • Op-amp filters
  • Soldering