Blog
Practical notes on LLM systems, evaluation, agents, inference, and developer workflows.
A model that picks from allowed actions can skip the essay. See what Jev actually discloses, how open Kev and SemIf score choices, and why calibration decides whether the result belongs in production.
Inside LeetLLM's Git-backed curriculum: reviewable article bundles, generated figures, user-owned state, and the checks between a source edit and a deployed page.
Compare AI coding subscription prices, annual commitments, seat minimums, usage windows, and overflow costs, then calculate a budget from reviewed work.
Set up a local Qwen3.6 llama.cpp server with a verified Unsloth GGUF. Check memory requirements and test text generation before adding MTP or vision.
Five focused AI engineering projects, with concrete experiments, inspectable artifacts, and failure cases that make the work worth discussing.
Start with programming, grow one useful project, and choose deeper study from the work you want to do. A practical path without a hiring countdown.
Read DeepSeek V4's checkpoints and benchmark conditions, then account for the V4.1 Flash API replacement before comparing deployment costs.
Choose an OpenClaw plan by permitted workload, credentials, quotas, and paid fallback. Compare dated prices and calculate request costs without confusing them with task costs.
Run Gemma 4 locally with Ollama. Choose a model tag, budget memory for context, verify GPU placement with ollama ps, and test the native API.
Compare vLLM, SGLang, TensorRT-LLM, Ollama, and llama.cpp by model support, latency, throughput, and cost. Choose a runtime for local or production use.
Fifty LLM engineering questions that connect model mechanics to serving, retrieval, agents, evaluation, and safety through evidence and failure diagnosis.
Compare Cursor, Codex, Copilot, and Claude Code through one authorization change. Choose by operating mode, review surface, cost meter, and team controls.
Compare AI engineering pay without mixing BLS wages, posted salary, and total compensation. Match role family, level, location, equity terms, and source date first.
Separate model rights from deployment control, evaluate the complete data path, and compare cost per accepted answer rather than API tokens against GPU rent.
Million-token windows can hold a bounded corpus, but capacity doesn't guarantee reliable retrieval. Measure effective context, watch length-based prices, and choose between full context and RAG.
Build a Python agent that follows configuration references, then test its tool permissions, history, recovery, and stop conditions before connecting a live model.
A policy assistant gives the wrong answer. Should you rewrite its prompt, retrieve better evidence, or train it? A controlled evidence swap reveals which experiment to run next.
Follow a support assistant from a wrong answer to a tested fix. See how applied AI, ML platforms, and research engineering divide the work.
Read SWE-bench as a repository-repair experiment: trace a patch through grading, compare variants and attempt rules, and check the limits of public scores. Updated for Multimodal v2.
Prepare for the interview you're actually taking: confirm the format, attempt representative problems, and use concrete coding, ML, and LLM debugging exercises to find gaps.
