Build notes from llm, embedding, AI, robotics and Linux servers

I tinker with embedded systems, run local AI models, wire up mini robots and keep a small rack of Linux servers alive. This is where I write down what worked, what broke and what I would do differently.

Latest

How to read language-model evaluation scoresA plain-English guide to held-out loss, nats, bits per byte, perplexity, and the PG-19 evaluation numbers from ghedi-c.llmevaluationlanguage-modelsGpu Unlock My RTX 3090 was doing tensor math at half speedGpu unlock results, NVIDIA caps FP32-accumulate tensor math on GeForce cards. I removed the cap on a 3090, measured it, and got 3 percent instead of the 60 the box promises.gpunvidiacudaBuilding an LLM from scratch15 partsA dependency-free GPT written in C — tokens, attention, and hand-written CUDA kernels, phase by phase, with the honest numbers that did not win.Latest — Part 15, Real prose, and what is left to build · November 2026The first jump: ESP32-S3 startup and memory, simplyA short guide to the ESP32-S3 memory used by a bare-metal program: where its code, data, and stack go, and why two addresses can name the same RAM.esp32s3baremetalxtensaESP32-S3 start.S and linker.ld, line by lineA guided reading of the startup assembly and linker script used by the ESP32-S3 bare-metal example, with each line tied back to the memory map.esp32s3baremetalxtensaEight threads on an 82-SM GPU: my first kernelsSix small CUDA kernels, from hello world to a tiled matrix multiply, and the Nsight report that showed how little of the RTX 3090 I was actually using.cudagpunvidiaThuruppu: An Agentic Card Game4 partsBuilding AI players for Twenty-eight, reusing prompt prefixes, serving with SGLang and vLLM, and tracing decisions while scraping inference metrics.Latest — Part 4, Tracing the game, scraping the model · February 2026Ray on Kind: GPU plumbing is the hard partA local GPU-enabled Ray cluster crosses several boundaries: host devices, Kind nodes, Kubernetes scheduling, KubeRay, and model-cache storage.kuberneteskindrayHome lab server13 partsRunning the workstation that doubles as the home server: the two-GPU split, Ubuntu, Docker, Pi-hole and the monitoring.Latest — Part 13, Watching the server without drowning in it · December 2025ATmega328P from scratch12 partsFlashing the chip directly: pins, interrupts, timers, buses and sensors, with no Arduino library in the way.Latest — Part 12, DHT11: a protocol built out of delays · May 2025Wake on Lan Proxy server for triggering remote wakesA small Rust proxy on a Raspberry Pi that wakes the home music server from a phone, and the magic packet, the signed cookie and the JSON field that silently did nothing.rustactix-webwake-on-lanSeven logic gates on a breadboard, and the one wired differentlyWorking through the 74HC family one chip at a time, from NOT to NAND, and the NOR chip whose pinout is not like the others.logic-gates74hcembeddedKernel modules: no main(), and a thread bug that hung the loadWriting a first Linux kernel module, adding parameters and kthreads, and the mistakes that taught me how the kernel expects a driver to behave.kernellinuxc

Archive: big data era (2015–2021)

22 posts from the big-data years: Spark, Kafka, Cassandra, Airflow and the rest of that stack.

Browse the archive →