Work
Tags
[1]
SeriesLLM Interpretability, Layer by Layer. A beginner-friendly tour of mechanistic interpretability on GPT-2 small, climbing a ladder of evidence: reading the residual stream, probes, sparse autoencoders, activation patching, circuits, steering, and finally a blind audit of a model with a hidden behaviour. Updated November 7, 2026.
[2]
SeriesConformal Prediction, Step by Step. A beginner-friendly guide to prediction intervals and sets you can trust: split and full conformal prediction, adaptive intervals, prediction sets for classifiers, distribution shift and time series, and conformal methods for LLMs, with the coverage guarantee checked at every step. Updated November 7, 2026.
[3]
SeriesTime Series, Era by Era. A beginner-friendly tour of forecasting, from moving averages and Holt–Winters through ARIMA, machine learning and deep learning to foundation models, all on one dataset, ending with an honest showdown. Updated November 5, 2026.
[4]
PostDo Your Annotators Agree? Cohen's Kappa, Krippendorff's Alpha and Friends. Why raw percent agreement misleads, how chance-corrected metrics fix it, when kappa breaks (the prevalence and bias paradoxes), and how to pick an agreement metric for human or LLM-as-a-judge labels. November 4, 2026.
[5]
PostThe Bradley–Terry–Luce Model: Turning Head-to-Head Wins into Scores. How one small formula turns pairwise wins into a ranking: the model, fitting it by maximum likelihood, when it fails to exist, and how it powers Elo, chatbot leaderboards and RLHF reward models. November 3, 2026.
[6]
PostLocal MLOps Mastery: Your Complete Guide to Building ML Systems on Your Machine. A practical 8-week roadmap for mastering the full MLOps lifecycle—data versioning, experiment tracking, model APIs, monitoring, and CI/CD pipelines—using only free, local, open-source tools. June 3, 2025.
[7]
PostBias-Variance Tradeoff in Machine Learning. A rigorous mathematical decomposition of prediction error into bias, variance, and irreducible noise—with practical intuitions on how to balance them for better-generalizing models. May 3, 2025.
[8]
PostUnderstanding AUROC: A Mathematical and Practical Perspective. Dive into the AUROC metric—its mathematical foundation, interpretation, and practical pros and cons in evaluating binary classification models. May 3, 2025.
[9]
PostScaling AI Rewards: How DeepSeek-GRM Redefines Generalist Reward Modeling. DeepSeek-GRM introduces Self-Principled Critique Tuning (SPCT) and inference-time scaling to create flexible, accurate reward models that rival much larger LLMs in generalist evaluation tasks. April 13, 2025.
[10]
PostDo Reasoning Models Really Say What They Think? A study exploring the faithfulness of chain-of-thought (CoT) in reasoning models. April 12, 2025.
[11]
ProjectOpen-RS: Enhancing Reasoning in Small LLMs with Reinforcement Learning. An open-source project exploring reinforcement learning techniques to boost reasoning capabilities in small large language models. (Python, PyTorch, vLLM, Hugging Face). April 3, 2025.
[12]
ProjectVITS-Japanese: Text-to-Speech for Japanese. An implementation of VITS for Japanese text-to-speech, enabling single-stage training and parallel sampling. (Python, PyTorch, NumPy, Torch). April 2, 2025.
[13]
ProjectHomemade Machine Learning: A Hands-On Journey into ML Algorithms. An educational project featuring Jupyter notebooks to implement and understand machine learning algorithms from scratch. (Python, Jupyter, NumPy, Pandas, Scikit-learn, Matplotlib). April 2, 2025.
[14]
ProjectNeurIPS LLM Efficiency Challenge 2023. Top 3 finish in NeurIPS 2023 LLM Efficiency Challenge, developing highly efficient LLMs under strict computational constraints. (Python, PyTorch, QLoRA, Transformers, PEFT). November 15, 2023.
[15]
ProjectGrapheme-to-Phoneme (G2P) Conversion. Sequence-to-sequence model with attention mechanism for converting written text to phonetic representations. (Python, PyTorch, Sequence-to-Sequence, Attention Mechanism). June 1, 2022.
