Latest Blog Posts

• May 7, 2026 • 20 min read Policy Gradient Methods: REINFORCE, Actor-Critic, and the Policy Gradient Theorem (S&B Ch. 13)
  

  Notes on policy approximation and its advantages, the policy gradient theorem and its proof, REINFORCE (Monte Carlo policy gradient), REINFORCE with baseline, actor-critic methods, policy gradient for continuing problems, and policy parametrization for continuous actions.

• Apr 17, 2026 • 3 min read Transformer Architecture Explained: Self-Attention, Encoders, and Decoders
  

  Quick overview of the Transformer architecture covering tokenization, attention, multi-head attention, encoders, decoders, masked attention, and cross attention.

• Apr 17, 2026 • 14 min read SAM 2 Explained: Meta's Promptable Visual Segmentation Model
  

  Quick overview on SAM 2, a foundation model for promptable visual segmentation in images and videos, covering the PVS task, model architecture, data engine, SA-V dataset, zero-shot experiments, and conclusions.

• Apr 4, 2026 • 13 min read Muon Optimizer Explained: Newton-Schulz Orthogonalization Beyond Adam
  

  From Muon optimizer to MuonClip covering Adam’s limitations, momentum 2D matrix, SVD-based orthogonalization, odd polynomial approximation, Newton-Schulz 5 iteration, the exploding attention logit crisis, and QK-Clip for MHA & MLA.

• Mar 16, 2026 • 4 min read Inkcast: Turn Any EPUB or PDF into an Audiobook in Your Browser
  

  A privacy-first, text-to-speech reader for EPUBs, PDFs, and web articles. Built in a single HTML file, no backend required.

• Mar 13, 2026 • 33 min read Eligibility Traces Explained: TD(λ), Sarsa(λ), and the λ-Return (S&B Ch. 12)
  

  Notes on the lambda-return and its variants - TD, n-step truncated, online, true online, Sarsa, Watkins’s Q, Tree-Backup - dutch traces in MC learning, variable bootstrapping and discounting, off-policy traces with control variates, stable off-policy methods with traces, and implementation issues.

• Mar 9, 2026 • 22 min read The Deadly Triad in RL: Off-Policy Learning with Function Approximation (S&B Ch. 11)
  

  Notes on semi-gradient off-policy methods, examples of divergence, the deadly triad, linear value-function geometry, gradient descent in the Bellman error, learnability, Gradient-TD methods, Emphatic-TD, and reducing variance.

• Mar 9, 2026 • 8 min read Semi-Gradient Sarsa and the Average Reward Setting in RL (S&B Ch. 10)
  

  Notes on episodic semi-gradient control, n-step Sarsa, average reward for continuing tasks, deprecating the discounted setting, and differential semi-gradient n-step Sarsa.

• Mar 4, 2026 • 10 min read Why AI Voice Assistants Fail Tonal Languages: ASR Bias in Igbo
  

  A systematic evaluation of tonal fidelity in facebook/omniASR-CTC-1B reveals 75% diacritic loss on Igbo language tone marks. Evidence suggests the model generates tones probabilistically rather than acoustically.

• Mar 1, 2026 • 13 min read Evaluating Tone Preservation in ASR: How Meta's omniASR Handles Igbo Diacritics
  

  Controlled evaluation of omniASR-CTC-1B on Igbo language using 21 systematically designed samples. Diacritic Error Rate (DER) quantifies tone-specific failures. Bootstrap resampling (10,000 iterations) yields 75.5% loss (95% CI [57.1%, 89.7%]). Monotone control reveals orthographic bias. Includes full methodology, code, and reproducibility guide.

• Feb 27, 2026 • 26 min read Value Function Approximation in RL: Tile Coding, Fourier Basis, and SGD (S&B Ch. 9)
  

  Notes on value-function approximation, the prediction objective (VE), stochastic gradient descent, linear methods, feature construction, ANNs, LSTD, memory-based and kernel-based approximation, and interest & emphasis.

• Feb 24, 2026 • 30 min read Model-Based RL Explained: Dyna-Q, MCTS, and Prioritized Sweeping (S&B Ch. 8)
  

  Notes on model-based RL, Dyna-Q, prioritized sweeping, trajectory sampling, RTDP, heuristic search, rollout algorithms, and Monte Carlo Tree Search.

• Feb 9, 2026 • 20 min read Machine Learning Math: From Linear Algebra to Attention — A Rigorous Deep Dive
  

  A rigorous exploration of machine learning mathematics, from information theory and linear algebra to optimization dynamics and generative models, with correct implementations and theoretical connections.

• Jan 22, 2026 • 17 min read Neural Network Optimizers Compared: SGD, Adam, AdamW, and Muon
  

  From SGD to Muon, explore how each optimizer builds on its predecessors to solve gradient descent challenges in deep learning.

• May 30, 2025 • 17 min read Machine Learning Equations Cheat Sheet: From Entropy to Attention
  

  A reference guide to fundamental machine learning equations with correct implementations and clear explanations of how they connect to each other.

• May 30, 2025 • 10 min read Backpropagation from Scratch: Chain Rule, Math, and Python Code
  

  A highly technical yet understandable exploration of backpropagation, detailing its mechanics, mathematical foundations, and practical applications, making it accessible to those with a basic understanding of calculus and machine learning.

• Feb 24, 2025 • 7 min read America's Energy Strategy After COVID-19: Natural Gas, Nuclear, and Renewables Reassessed
  

  A post-COVID-19 critical review of Robert Bryce’s “Power Hungry: The Myths of 'Green' Energy and the Real Fuels of the Future.”

• Aug 2, 2024 • 9 min read Building an RNN from Scratch in Python: Character-Level Language Model with BPTT
  

  In this blog post, we’ll dive deep into the implementation of a character-level language model using a vanilla Recurrent Neural Network (RNN). This type of model can learn to generate text one character at a time, capturing the patterns and structure of the language it’s trained on. We’ll walk through…

• Jul 15, 2024 • 3 min read Karpathy's Neural Networks Zero to Hero: Complete Implementation Notes
  

  This blog post presents my detailed implementation of Andrej Karpathy’s Neural Networks: Zero to Hero YouTube lecture series and exercises in Jupyter Notebook. The articles go deeply from NNs to MLPs to CNNs to LLMs to ensure a thorough and robust understanding of neural networks.

• Sep 20, 2020 • 10 min read Why Natural Gas to Nuclear (N2N) Can't Solve America's Energy Problem: A Critical Review
  

  A pre-COVID-19 critical review of Robert Bryce’s “Power Hungry: The Myths of 'Green' Energy and the Real Fuels of the Future.”