Deep learning tutorials, neural networks, computer vision, and AI implementation guides
Reading the Transformer paper in 2024 with production LLM experience is a different exercise than reading it in 2017. What the paper got right, what it underspecified.
Same YOLO v8 model, three hardware targets, wildly different power/latency profiles. Numbers from an afternoon of benchmarking on a controlled workload.
Retrieval-augmented generation and fine-tuning solve different problems. Using the wrong one wastes money and produces confusing failures. A decision framework from production.
JSON mode and function calling replaced most of my prompt gymnastics. What I spent time optimizing before those features existed and what still requires careful prompting.
text-embedding-ada-002 is not always the right choice. Benchmarking five open-weight models on our domain-specific retrieval task produced a surprising ranking.
Post-training quantization to INT8 worked beautifully on the benchmark — and lost 4 points on our production distribution. An anatomy of the failure.
Comprehensive guide to implementing MLOps practices for scalable, maintainable machine learning systems in production environments with automated training, deployment, and monitoring.
Feature distribution drift is not the same as label drift and neither is the same as business metric degradation. Confusing them leads to alert fatigue.
Missing hyperparameters, unreported data cleaning steps, and hardware-dependent results make ML reproducibility harder than most papers admit.
