Machine Learning Engineer at Meta, Kaggle Grandmaster, and Google Developer Expert writing about applied ML, AI paper reviews, systems, and learning.
Kimi K3 is a 2.8-trillion-parameter open-weight multimodal MoE model that combines hybrid linear–global attention, selective residual connections across depth, and large-scale agentic reinforcement learning.
A review of the second edition of Python for Algorithmic Trading Cookbook by Jason Strimpel, a 68-recipe book that goes from getting raw data to live execution on the Interactive Brokers API.
Harness Handbook proposes a new way to navigate large AI agent codebases by organizing them around behaviors instead of files. The result is better code localization, higher-quality edit plans, and more efficient agent engineering.
A review of Time Series with PyTorch, a broad deep-learning-for-forecasting book that teaches from honest evaluation and PyTorch fundamentals through RNNs, transformers, conformal prediction and diffusion, and on into classification, clustering, and anomaly detection.
I used a Claude Code starter kit and Google Antigravity with Gemini 3.1 to participate in a Kaggle Playground competition. Most of the work was catching the silent bugs and stopping the agent from taking wrong turns.
Two experiences from two CayleyPy Kaggle competitions: how learned-heuristic beam search solves enormous Cayley graphs, and how Claude Code simplifies and speeds up the engineering loop.
A review of GPU-Accelerated Computing with Python 3 and CUDA by Niels Cautaerts and Hossein Ghorbanfekr, a hands-on book that goes from raw Numba-CUDA kernels through CuPy, RAPIDS, and JAX to four full applications, with a profile-everything approach.
A review of LocateAnything, an NVIDIA vision-language model that treats each bounding box as one atomic unit and decodes it in a single parallel step instead of a sequence of coordinate tokens. Its Parallel Box Decoding reaches roughly 2.5x the throughput of the nearest grounding VLM while improving high-IoU localization.
MiniMax Sparse Attention is a practical sparse-attention design for million-token LLMs - it uses a lightweight learned indexer to select relevant KV blocks and performs exact attention only over those blocks. The paper is important because it connects architecture, training stability, and GPU kernels into a deployable long-context system, powering the open-weight MiniMax-M3 model.
A hands-on look at MiniMax M3 through Claude Code — what its new MiniMax Sparse Attention (MSA) is and how it differs from the lightning-attention and full-attention designs of earlier MiniMax models, plus three real tasks: auditing and refactoring an old idle game, debugging two stubborn UI bugs from screenshots, and turning years of Spotify history into music recommendations.