A datetime.now() at the head of an agent’s system prompt was re-rendered every turn, so the provider’s prompt cache missed on nearly every request. Prefix caching is exact-match and silent, so prefix poisoning costs money without ever failing a test. Covers the lying cost-per-run metric, the render-once fix that beats freezing the timestamp, and how to measure the cache-hit ratio lift honestly (32.9% to ~53%).

Why writing culture matters more in the AI era: LLM agents are “literal readers” that do exactly what your docs say and fill in none of the gaps a human silently patches. Built on a shared STATUS.md, the case for writing things down, and where it tips into over-documenting.

A working read of the EU AI Act for a small B2B startup building on a foundation model: provider versus deployer, the four risk tiers, and why the Aug 2, 2026 transparency deadline (Article 50) is a disclosure task, not a rebuild. Plus the Digital Omnibus delaying high-risk obligations to Dec 2027, the AI-literacy duty, and the open question of who owes end-user disclosure down a B2B chain.

A 1M-token LLM context window is a tool, not a target. How context actually works, why long threads cost more on every turn, and when to start a fresh chat versus keep going, with practical Claude Code tips and /context.

A file-based pattern for coordinating multiple LLM agents on one task: a single shared, per-feature STATUS.md the agents read and write under explicit rules, how to make them follow a protocol and survive context loss, and when not to use it.

What an optocoupler is and how electrical isolation works, built from a discrete LED and phototransistor on a breadboard with measurements, then replaced with a real 4N35 chip, and where optocouplers show up in actual circuits.

Building a DIY Bluetooth page-turner remote for a Kobo e-reader by making an ESP32-S3 Lilka device act as a BLE HID keyboard, including BLE debugging on Debian, testing on the Kobo Clara 2E, and the full project code.

Building a minimal LLM Ops stack: a FastAPI “customer support reply drafter” instrumented with Langfuse for request tracing, grounded retrieval, and per-request model cost tracking, so every LLM call is inspectable.

Retrieval-Augmented Generation (RAG) explained without buzzwords: how it gives an LLM the right data at answer time to fix stale knowledge and hallucinations, the step-by-step flow, its benefits over fine-tuning, and when RAG is not the answer.