Software engineering, AI systems and technical paper reviews.
I built a FastAPI service that turns a two-step, rate-limited carrier API into one internal endpoint — token caching, retry logic, order mapping, and an undocumented limit found the hard way.
A build log for a RAG system over thousands of real drug leaflets — PDF-to-markdown, section-aware chunking, a cross-encoder reranker, and GPU-scaled embeddings. The bugs that nearly sank it, with the numbers.
How I predicted California house prices with linear regression trained by gradient descent (SGDRegressor) — mutual-information feature selection, a scaling gotcha that breaks SGD, and an honest benchmark against KNN, a decision tree, and WEKA.
How I built a speech emotion recognition model in Python — parsing MFCC voice features, pruning them with mutual information, and benchmarking five classifiers (logistic regression, naïve Bayes, KNN, decision tree, MLP) in scikit-learn against WEKA.
A practical, end-to-end guide to building a private RAG system that never leaves your machine — local embeddings, vector store, and LLM, with real code.
I write about the backend×AI wedge — wins and failures included. Follow along, or say hi.