Clear, reusable notes on mathematics, statistics, machine learning, NLP, and LLMs—connected to real experiments in evaluation, multilingual AI, retrieval, and applied ML.
Understand the ideas
Mathematics · Statistics · ML · NLP & LLMs
Test them with evidence
Questions · Experiments · Evaluation
Make them work
Multilingual AI · Retrieval · Applied ML
Foundations
This track is where existing high-quality resources become my own explanations, notes, examples, and interview-ready mental models.
Linear algebra, calculus, optimization, graphs, and representation spaces explained through ML and NLP examples.
Probability, distributions, confidence intervals, bootstrap testing, hypothesis tests, and uncertainty in evaluation.
Splits, generalization, model selection, cross-validation, losses, regularization, and practical training workflows.
Tokenization, embeddings, Transformers, fine-tuning, prompting, sequence models, and language-specific evaluation.
Experiment repositories, config-driven pipelines, logging, annotation sheets, demos, and reproducible analysis.
Research & Applications
This track avoids boxing the blog into one thesis topic. It groups research by capability: building, evaluating, explaining, retrieving, and applying AI systems.
Machine translation, multilingual NLP, low-resource adaptation, terminology, domain adaptation, and language variation.
Metrics, benchmark design, statistical testing, ablations, error analysis, human evaluation, and robustness.
Interpretability, bias analysis, attribution, model auditing, SHAP/LIME, Shapley values, and controlled evaluation.
Information retrieval, RAG, long documents, graph retrieval, evidence grounding, and document intelligence.
Data pipelines, classification, forecasting, information extraction, deployment, product demos, and applied analytics.
Foundations posts build reusable mental models; research posts connect them to experiments, evidence, and real systems. Chinese translations are available from the language switch on selected articles.
A visual guide to multi-hop evidence, graph-based retrieval, query planning, iterative search, verification, retry, and the boundaries of Agentic RAG.
A three-layer framework for RAG evaluation: whether retrieval found the right material, whether the evidence is sufficient and complete, and whether the answer is grounded and correctly cited.
A visual guide to candidate generation, BM25 and dense fusion, Reciprocal Rank Fusion, cross-encoder reranking, and the limits of a two-stage retrieval pipeline.
A visual comparison of inverted indexes, BM25, embeddings, dense retrieval, and the failure modes that determine which evidence a RAG system can find.
A research story about turning 15 years of annual-report PDFs into an auditable benchmark for retrieval, GraphRAG, failure recovery, and grounded question answering.
A visual guide to RAG document preparation: PDF parsing, layout and tables, evidence objects, chunking, metadata, embeddings, and indexes.
Portfolio pages show what I built. These notes explain the foundations, decisions, experiments, and tradeoffs behind the work.