Joe Ma's Tech BlogArticles, showcases, and experiments by Joe Ma.https://mrjoema.github.io/Compressing a ResNet for the edge: a cat breed + head-box detectorhttps://mrjoema.github.io/blog/cat-breed-detector/https://mrjoema.github.io/blog/cat-breed-detector/An end-to-end model-compression project: fine-tune a multi-task ResNet50 to classify cat breeds and regress a head bounding box, then prune and quantize it down to a 19 MB int8 model at ~0.90 accuracy — a 5× size cut. With the corrections I made along the way.Thu, 25 Jun 2026 00:00:00 GMT96% full-board accuracy on chess recognition — what the papers say and what actually workedhttps://mrjoema.github.io/blog/chess-recognition/https://mrjoema.github.io/blog/chess-recognition/A results-first walkthrough of building a chess position recognizer: 99.85% per-square accuracy in 30 epochs, backed by five papers. Every design decision has a citation.Wed, 27 May 2026 00:00:00 GMTCNN from 0 — why convolutional networks exist and what every layer actually doeshttps://mrjoema.github.io/blog/cnn-from-zero/https://mrjoema.github.io/blog/cnn-from-zero/A from-scratch tour of convolutional neural networks: why they replaced manual feature engineering, what every layer does, how the network actually learns (softmax, cross-entropy, backprop), and how it assembles into a working VGG. Inline diagrams and the math behind each layer.Sun, 31 May 2026 00:00:00 GMTWhat continuous batching changes: reading three LLM inference traceshttps://mrjoema.github.io/blog/llm-batching-trace-comparison/https://mrjoema.github.io/blog/llm-batching-trace-comparison/A trace-first comparison of serial inference, static batching, and continuous batching: model-call count, batch shrinkage, refill, and the latency-throughput connection.Thu, 30 Jul 2026 00:00:00 GMTWhere does LLM inference time go? Prefill, decode, and KV cache on Apple siliconhttps://mrjoema.github.io/blog/llm-prefill-decode-apple-silicon/https://mrjoema.github.io/blog/llm-prefill-decode-apple-silicon/A ground-up implementation and profiling study of autoregressive inference on an M4 Pro: why prefill and decode stress the GPU differently, how KV cache grows, and how the MPS/Metal path differs from CUDA.Sun, 19 Jul 2026 00:00:00 GMTNeural network pruning — from fine-grained to channel pruning and everything in betweenhttps://mrjoema.github.io/blog/nn-pruning-deep-dive/https://mrjoema.github.io/blog/nn-pruning-deep-dive/A lab-driven exploration of neural network pruning: implementing fine-grained and channel pruning, sensitivity analysis, channel sorting, and the hardware trade-offs of sparse vs. structured weights.Mon, 08 Jun 2026 00:00:00 GMTPipe vs shared memory: what actually happens when you pass a tensor between processeshttps://mrjoema.github.io/blog/pipe-vs-sharedmem/https://mrjoema.github.io/blog/pipe-vs-sharedmem/A benchmark-driven deep dive into OS primitives — virtual memory, kernel buffers, and why shared memory is 2.8× faster than pipe + pickle for passing tensors between Python processes.Wed, 27 May 2026 00:00:00 GMTNeural network quantization — from k-means codebooks to integer-only inferencehttps://mrjoema.github.io/blog/quantization-deep-dive/https://mrjoema.github.io/blog/quantization-deep-dive/A lab-driven exploration of neural network quantization: k-means codebooks, linear (affine) quantization with scale and zero-point, integer-only inference, and the PTQ-vs-QAT trade-offs that decide how much accuracy survives on the edge.Tue, 16 Jun 2026 00:00:00 GMT