Case study
Neural networks written from scratch in NumPy: a backpropagation MLP for flood forecasting and classification, and an MLP trained by a genetic algorithm that reaches 97.5 % accuracy on breast-cancer data.
Course assignments where no machine-learning library was allowed for the network. Assignment 1 is a multilayer perceptron with backpropagation and momentum, used for regression (forecasting a river level 7 hours ahead from two stations) and classification (a two-class pattern set). Assignment 3 replaces backpropagation with a genetic algorithm that evolves all the weights of an MLP to classify breast tumours in the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. Fuzzy logic was covered in the same course.
97.5 %
WDBC accuracy, GA-trained MLP with 32 hidden nodes (±2.1 %, 10-fold CV)
0.91
R² forecasting the water level 7 hours ahead (MSE 0.0075)
94.4 %
Cross-dataset CV accuracy; 20/20 correct on the held-out test set
Library training hides how weights are updated, how data is split, and how a search behaves. Writing every step myself makes those mechanics visible and testable.
An academic study of model behaviour, not a clinical diagnostic tool.
Individual implementer — models, training loops, experiments, and reports (CPE 261456 Introduction to Computational Intelligence).
Neural networks
Evolutionary search
Experiment design
Writing the MLP directly trades speed and convenience for transparency; every update rule had to be written and debugged by hand.
10-fold cross-validation reduces dependence on one lucky train/test cut. It still does not prove the model would work on a new clinical population.
AI / ML
LiveSpeak to your phone and your Windows PC does it — with no cloud AI. Speech-to-text and an LLM run on my HomeLab GPU; a policy layer and an HMAC-signed agent decide what is allowed to run.
Infrastructure
An old PC rebuilt as a Proxmox server that runs my live projects and a local LLM on its own GPU, published through Cloudflare Tunnel with Grafana and Dozzle for metrics and logs.
