KV-cache quantization benchmark Comparing FP16, INT8, and INT4 KV-cache compression on Llama 3.2-1B with HQQ — memory, throughput, and quality trade-offs. GPT from scratch A character-level GPT built from scratch in PyTorch — token embeddings, multi-head self-attention, feed-forward blocks, trained on tiny Shakespeare. Multimodal AI search engine Semantic search across an image corpus using CLIP embeddings and FAISS — supports text-to-image and image-to-image queries with sub-second response times. Bank term-deposit classification Binary classification on a Kaggle bank-marketing dataset — Logistic Regression, XGBoost, LightGBM, and a stacked ensemble, evaluated with 5-fold stratified ROC-AUC. Sudoku solver Browser-based Sudoku puzzle solver — interactive grid in JavaScript with a backtracking solver under the hood. To-do list app A full-stack to-do app with Node.js + Express, EJS templates, MongoDB via Mongoose, and Bootstrap on the frontend. Early CRUD project from when I was learning the stack.