I work across software systems, machine learning, and product implementation — turning coursework and operational problems into working software while documenting the decisions behind it.
Implements CNN image classification and image captioning encoder-decoder from scratch using only NumPy. Includes LocallyConnected2D, LSTM cells, greedy decode, beam search, Grad-CAM, and backward propagation via tensor autograd.
Featured case studies
Work with a story behind it.
Scroll the narrative. The project preview starts in the center, then docks into the context rail as the selected work becomes the focus.
01 / Python
Overfitcheck - CNN & RNN from Scratch
Implements CNN image classification and image captioning encoder-decoder from scratch using only NumPy. Includes LocallyConnected2D, LSTM cells, greedy decode, beam search, Grad-CAM, and backward propagation via tensor autograd.
Role
Individual implementation
Owned
Implemented the NumPy tensor autograd path, CNN layers, LSTM cells, decoding, and Grad-CAM tooling.
Approach
Kept the forward and backward passes explicit so the learning mechanics remain inspectable instead of hiding them behind a framework.
Hardest decision
Used one small tensor/autograd core for both image classification and image captioning.
Result
Two model tracks, two decoding modes, and an interpretable Grad-CAM output are documented in a runnable repository.
Implements Decision Tree (ID3, C4.5, CART), Logistic Regression with SGD, and SVM (OvA, OvO, DAGSVM) from scratch in Python for student dropout prediction classification.
Role
Individual implementation
Owned
Implemented the Decision Tree variants, SGD logistic regression, and SVM multiclass strategies from scratch.
Approach
Compared tree-based and margin-based classifiers through a shared student-dropout prediction task.
Hardest decision
Kept ID3, C4.5, CART, OvA, OvO, and DAGSVM as explicit implementations so their trade-offs can be compared.
Result
Three tree algorithms, one logistic model, and three SVM decision strategies are available in the repository.
CV-based applicant tracking system using KMP, Boyer-Moore, and Aho-Corasick pattern matching with Levenshtein distance for keyword similarity. Parses PDF CVs and surfaces ranked matches for recruiters.
Role
Individual implementation
Owned
Built the PDF parsing, keyword matching, similarity scoring, and recruiter-facing PyQt flow.
Approach
Combined KMP, Boyer-Moore, Aho-Corasick, and Levenshtein distance so exact and fuzzy keyword evidence can be ranked together.
Hardest decision
Used multiple pattern-matching algorithms as explicit interchangeable strategies instead of one opaque search implementation.
Result
Three exact-pattern algorithms plus fuzzy similarity produce ranked CV matches from parsed PDF input.
19 projects live in the archive, from compilers and search systems to ML models built from scratch. Each one is documented so the decisions stay visible.
