Selected Work · Engineering Systems
Featured Projects
Architectures in computer vision, parameter-efficient language models, and foundational deep learning.
01Computer Vision · Deep Learning
View RepositorySatellite Semantic Segmentation
Encoder-decoder segmentation pipeline featuring a custom boundary fusion module engineered for pixel-level classification of high-resolution satellite imagery.
Mean IoU
84.6%
Dice Score
89.2%
Evaluation Datasets
LandCoverNet & DeepGlobe
Key Technical Contributions
- Engineered multi-scale semantic feature fusion coupled with boundary information to preserve fine-grained land cover details and minimize edge distortion.
- Benchmarked against standard UNet baselines, demonstrating significant spatial delineation improvements across fragmented terrain classes.
Stack:PyTorchComputer VisionSegmentationBoundary FusionDeepGlobe
02Natural Language Processing · PEFT
View RepositoryQwen LLM Fine-Tuning
Parameter-efficient supervised fine-tuning of the Qwen2.5-1.5B Instruct language model utilizing Low-Rank Adaptation (LoRA) and the Unsloth acceleration kernel.
Base Model
Qwen2.5-1.5B Instruct
Adapter Config
r = 16, α = 32, dropout = 0.05
Kernel Framework
Unsloth & PyTorch
Key Technical Contributions
- Curated task-specific instruction prompt-response pairs to specialize reasoning behavior while eliminating memory bottlenecks during backpropagation.
- Evaluated response coherence and loss trajectory across held-out test splits, verifying domain adaptation fidelity without catastrophic forgetting.
Stack:Qwen 2.5LoRAPEFTUnslothPyTorchInstruction Tuning
03Neural Architecture · Foundations
View RepositoryLanguage Model from Scratch
Character-level autoregressive neural language model constructed entirely from scratch in PyTorch without high-level library abstractions.
Corpus Size
93 Million Characters
Architecture
Embedding + Bigram Neural
Sampling Strategy
Autoregressive Softmax
Key Technical Contributions
- Implemented custom vocabulary tokenization, embedding parameterization, logit projection, and cross-entropy optimization pipelines.
- Engineered text generation via autoregressive probabilistic character sampling with tunable temperature scaling.
Stack:PyTorchTokenizationNeural NetworksCross-EntropyAutoregressive