A comprehensive collection of my work across various domains.
Wafer-map pattern recognition on real WM-811K data: lot-disjoint evaluation, a tested idea that failed, conformal prediction that holds on unseen lots, and an ONNX + FastAPI service.
GPT-style transformer trained on 134k QM9 molecules for inverse molecular design — target a HOMO/LUMO value, generate the SMILES that satisfies it.
GNN classifier that identifies thermodynamic phases (FCC, BCC, liquid, amorphous) from raw MD trajectories.
Fab yield excursion detection on a simulated fab with ground truth: SPC and root cause scored against it, FabEye wafer-map patterns, real-time SPC on Kafka, Grafana and Power BI.
In-line wafer-map watchdog at the equipment edge: zero-copy SECS-II decode, CPU/GPU spatial-signature detectors, FabEye's CNN on a custom CUDA inference engine, lot hold sent back to the tool in under a millisecond.
Step-by-step NN implementation without any deep learning frameworks.
ViT implementation exploring transformer applications in computer vision.
Hands-on coding of ResNet, AlexNet and other popular CNN architectures.
Tiny from-scratch automatic differentiation engine for deep learning fundamentals.
LSTM-based model to identify misinformation by analyzing textual data.
Predicting future prices using LSTM networks on historical market trends.
LSTM model for predicting equipment failures through time-series sensor analysis.
Deep learning classifier for German traffic signs (GTSRB dataset).
Collection of supervised and unsupervised learning techniques.
Python framework for molecular dynamics simulations using Langevin dynamics.
Simulating metal deformation using LAMMPS and post-processing via OVITO Python API.
Obstacle avoidance and path planning using Gazebo and ROS 2 Humble.
Velocity command navigation in Gazebo simulation via geometry_msgs/Twist.
Calibrated RGB detection using HW-531 sensor and Arduino Uno.
Gravity-bound particle simulation with collision physics in C++.
Dynamics of colliding circles using momentum and velocity principles.
Interactive visualization of BFS, DFS, A* and Dijkstra solving mazes.
Machine learning to predict match outcomes and analyse player performance.