From GGUF to GPU: How llama.cpp Actually Runs a Model
Trace llama_decode through GGML graphs, CUDA kernels, GPU memory, and next-token generation—and see why expensive LLM inference can still be fast.
Read article →The field journal
Practical Architecture Writing
Practical architecture decisions, tradeoffs, and operating lessons from software that has to survive business change.
Browse all topics →Find the right article
Showing 5 of 56 articles
Latest writing
More practical architecture writing
An LLM Is Not the Program: What Is Actually Inside the Model File
Follow an LLM request past the API and into GGUF tensors, learned parameters, attention, and the transformer layer that turns model data into logits.
Membership Reconciliation: When an Audit Stops at 10,000 Records
How a silent 10,000-record cap, ambiguous identities, and competing sources of truth changed a production membership reconciliation.
Build Around Business Capabilities, Not Vendor Features
Why durable business capabilities should define platform boundaries, while commerce, payment, and ERP vendors remain replaceable implementation details.
pg-boss in Production: Reliable Background Jobs for Node.js
What I learned running pg-boss in a high-volume Node.js backend: why PostgreSQL made the queue dependable, how PM2 protected web traffic, and why every job still needed to be safe to retry.