Built around model design, training, and the practical side of dataset preparation.
Model architecture, data collection, preprocessing, training, and evaluation.
Deep-learning experimentation with supporting OpenCV and TensorBoard tooling.
Problem Framing#
The project goal was to recognize signs from the LIS alphabet through a computer-vision pipeline. That makes it a model problem, but not only a model problem. Accuracy depends heavily on how data is collected, organized, and cleaned before training ever begins.
That is why this project matters to me more than a simple “trained a model” summary. The interesting work was in building the surrounding pipeline well enough that the experiments meant something.
Model Work#
I implemented both custom CNN architectures and AlexNet-based variants, comparing approaches rather than treating the first solution as final.
The practical goal was to understand the tradeoff space:
- how much complexity the task actually needed
- how the models behaved under different data assumptions
- what architecture decisions improved generalization instead of only fitting the training set
Dataset and Tooling#
Data acquisition and data engineering were a core part of the project. In machine-learning work, that stage often determines whether the modeling phase is credible or just decorative.
I built the pipeline around:
- dataset preparation for training and evaluation
- image handling and preprocessing support
- experiment visibility through tooling such as TensorBoard
That combination made it possible to iterate with feedback rather than treat training as a black box.
What This Project Shows#
This article belongs in the portfolio because it demonstrates comfort outside game-engine work. It shows that I can move into a different technical domain, reason about pipelines end to end, and structure experiments in a way that is useful rather than superficial.