// 03 — Projects
Things I've Built
VR · Data Visualization · Unity
Immersive US Census Data Visualization
A VR app on Meta Quest 3 that renders 3,100+ interactive 3D data points of US Census data (income, population, poverty) with dynamic labeling and gradient coloring by county.
Unity
Meta Quest 3
C#
OpenXR
XR Toolkit
NLP · LLM · Fine-tuning
Formal Speech → Gen Z Meme Language
Fine-tuned FLAN-T5-Base with QLoRA (4-bit) on 28k formal-to-slang pairs. Injected 210 custom emoji/slang tokens, slashing OOV rate from 18% to 5% and outperforming LSTM seq2seq baselines on BLEU & ROUGE-L.
FLAN-T5
QLoRA
HuggingFace
WandB
Google Colab
⚡ Google Colab · GPU T4
Time Series · Deep Learning
Time Series Forecasting (ETTh Dataset)
End-to-end oil temperature prediction system. Trained and compared LSTM, 1D CNN, and N-BEATS models — achieving a best test MAE of 0.76°C vs 1.63°C baseline, a 53% accuracy improvement.
N-BEATS
LSTM
1D CNN
PyTorch
Python
⚡ Google Colab
Computer Vision · Segmentation
Image Segmentation with U-Net
U-Net model with transfer learning (VGG16 encoder), data augmentation, and increased depth. Optimized MeanIoU via systematic ablation study comparing architectures.
U-Net
VGG16
TensorFlow
Keras
⚡ Google Colab
CNN · Classification · Transfer Learning
RPS Image Classification (CNN)
CNN classifier for Rock/Paper/Scissors images achieving ~99% validation accuracy using DenseNet121 transfer learning and extensive data augmentation.
DenseNet121
TensorFlow
Keras
Data Augmentation
⚡ Google Colab
ML · Game Design · Decision Trees
Guess the Animal (ML Game)
ML-based 20-questions-style game using the AWA2 dataset. Compared SVM, Neural Networks, and Random Forests — ultimately deploying a decision tree classifier for optimal interactive game performance.
Scikit-Learn
AWA2 Dataset
Decision Trees
Python
Web App · NLP · Audio Synthesis
Text-Based Video Synthesis
Web app converting text scripts into videos by integrating TTS, automated image retrieval, and Gentle forced aligner for word-level audio-visual synchronization. Output via FFmpeg pipeline.
Python
FFmpeg
TTS
Gentle Aligner
Web
NOTE ON COLAB PROJECTS
Several projects were developed in Google Colab notebooks. Screenshots and experiment reports (tracked via Weights & Biases) are available on request — feel free to reach out.
