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.
Deploying a new model version to 5% of traffic while monitoring both model-specific and business metrics is the right approach. The routing layer and rollback trigger that make it practical.
Building a gold evaluation dataset costs time you often don't have. Three evaluation approaches — LLM-as-judge, behavioral regression tests, and slice-based metrics — that work without one.
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.
pgvector keeps your stack simple if you already use Postgres. Qdrant's HNSW index and filtered search performance outpace pgvector at scale. The crossover point we measured.
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.
Whisper.cpp on a Raspberry Pi 5 achieves real-time transcription for short utterances. The latency budget at different model sizes and the memory footprint that limits deployment.
Comprehensive guide to implementing MLOps practices for scalable, maintainable machine learning systems in production environments with automated training, deployment, and monitoring.
TFLite Micro and ONNX Runtime both target microcontrollers. The supported operator sets, flash footprint, and inference speed on Cortex-M55 differ enough to matter for model selection.