Personal Blog

I spent time understanding the inference engine landscape and built my own reference. Here is what I learned about vLLM, TGI, TensorRT-LLM, llama.cpp, ONNX Runtime, and PyTorch native, and when each one earns its place.

Most people think more context means better output. Past a threshold, the opposite is true. The skill is curation: choosing what to include, what order to put it in, and what to deliberately leave out.

Hand the same AI assistant to two engineers and you get opposite results. The tool changes your speed, never your standard. The quality bar is set by the person using it, and that is still the hardest thing to build.

Every quality check I add to an AI pipeline is either a hard gate that stops the output from shipping, or it does not exist. Warnings that let you proceed are noise within a week.

People are wiring AI agents into production with permissions no human would get on day one. The fix isn't clever prompting. It's the same boring guardrails we use for new hires.

I stopped asking one AI to review everything at once. Six focused agents, each checking a single dimension, produce sharper feedback in a third of the time.

The same model that fabricates a claim will flag it as wrong five minutes later in a fresh session. I use that asymmetry as my primary workflow now.

I turned a recruiter's profile optimization guide into a scoring tool that reads your LinkedIn data export and tells you exactly what to fix.

I stopped writing long prompts months ago. Everything I do with AI is skill-based now, and the difference in consistency and speed is enormous.