Blog
Practical notes on LLM systems, evaluation, agents, inference, and developer workflows.
How LeetLLM turns research into curated lessons with research packets, article bundles, validation checks, generated diagrams, component-based illustrations, and a production web stack.
Run Qwen3.6 locally with llama.cpp and Unsloth GGUF. Pick between 27B dense and 35B-A3B MoE, choose a quant, expose a local endpoint, then test MTP after the plain run works.
Five portfolio projects that prove real AI engineering skill: shipped demos, eval reports, traces, cost notes, tests, and design docs.
A practical path from beginner to hire-ready AI engineer: programming basics, LLM APIs, RAG, evals, agents, deployment, and portfolio proof.
DeepSeek V4 pairs open weights, 1M context, and low hosted pricing with strong agentic coding claims. Use cheaper lanes when evals pass, and escalate when quality or policy requires it.
OpenClaw plan selection is a routing and quota problem. This July 2026 update keeps Fire Pass out of the buying list and compares MiniMax Token Plan, Qwen Cloud Coding Plan, Z.AI GLM Coding Plan, and OpenAI using current official docs.
Gemma 4 now has a 12B laptop lane, MTP drafters, and QAT checkpoints. Pick an Ollama tag, tune context, and keep local inference on the fast memory path.
Raw throughput is only half the inference-engine decision. Read an H100 benchmark snapshot, reason about KV-cache pressure, and choose between vLLM, SGLang, TensorRT-LLM, and Ollama.
Fifty LLM engineering concepts, organized by system layer. Each answer focuses on mechanism, trade-off, failure mode, and production intuition.
AI engineering pay is not one market. Use public job postings, Levels.fyi snapshots, and H-1B base records to benchmark offers by role family, level, location, and scope.
The real 2026 decision is whether you need control over weights and deployment, or the speed of a managed frontier API. Walk through a concrete software example, an illustrative cost model, and common mistakes that trip up teams.
Long-context windows help when relationships across a bounded corpus matter. Know what fits, what breaks, how to evaluate effective context length, and when the economics justify using it.
Build a working AI agent from the raw loop: define tools, let the model choose one, execute it in Python, append the observation, and add the guardrails that keep agents reliable.
Every LLM project starts with the same architecture question: improve the prompt, add retrieval, or fine-tune the model. Use a practical decision rule to choose the next layer.
AI engineering sits between foundation models and product engineering. The day-to-day work behind useful LLM systems is prompts, RAG, tools, evals, serving, and lightweight adaptation.
SWE-bench measures whether coding agents can fix real repository issues: task format, scoring, benchmark variants, contamination limits, and leaderboard claims.
A practical guide to ML and LLM engineering interview prep in 2026, covering classical ML filters, LLM systems design, evaluation, and a concrete study roadmap.
