AI

Design decisions for production LLM systems — routing, cost, guardrails, and multi-model orchestration. The layer between running models and building reliable AI applications.

LLM costs scale linearly with usage. A system processing 10,000 requests a day at $0.01 per request costs $100 daily — $365 a year. At enterprise scale, that’s over $10,000.

LLMs are unpredictable. They hallucinate, leak data, generate harmful content, or refuse legitimate requests. Guardrails constrain model behavior without sacrificing capability.

Running a 70B parameter model to summarize a 200-word email is wasteful. Running a 3B model to review production code is reckless. Most systems live somewhere in between — and that’s where model routing comes in.

Single-model systems are simple. Multi-model systems are powerful. The challenge isn’t choosing models — it’s designing the architecture that orchestrates them.

Memory turns assistants from reactive to persistent, but it is also where many systems quietly rot. Surveys argue the short-term versus long-term split is no longer enough for modern agent memory; OpenAI and LangGraph SDKs point to a simpler stack — working memory, durable state, and retrieval.

A production AI assistant is not “an LLM with a prompt”. It is a system that accepts intent, keeps state, decides when to retrieve or act, and exposes enough runtime detail to debug failures.

AI is not replacing knowledge management; it is changing the shape of it for both individuals and teams.

Open-source AI agent frameworks are exploding in popularity on GitHub. Two projects at the core of the self-hosted AI systems ecosystem — OpenClaw and Hermes Agent — have pulled so far ahead that the rest of the field is fighting for a distant third place.

I tested Speculative decoding (Multi-Token Prediction, MTP) performance in Qwen 3.6 27B and 35B on an RTX 4080 with 16 GB VRAM.

llama.cpp router mode is one of the most useful changes to llama-server in years. It finally gives local LLM operators something close to the model management experience people expect from Ollama, while keeping the raw performance and low-level control that make llama.cpp worth using in the first place.

The premise is simple: compiled knowledge is more reusable than retrieved fragments. RAG became the default answer to a straightforward question - how do I give an LLM access to external knowledge?

PKM, RAG, wikis, AI memory systems, and now practical AI-assisted workflows are often discussed as if they solve the same problem. They do not. They all deal with knowledge, but they operate at different layers:

Most LLM “structured output” tutorials are unserious. They teach you to ask for JSON politely and then hope the model behaves. That is not validation. That is optimism with braces.

This page is a practical reference for agentic LLM inference tuning (temperature, top_p, top_k, penalties, and how they interact in multi-step and tool-heavy workflows).

You already chat to Hermes Agent from your phone with text. Now you want to talk to it directly and get spoken replies back. That is usually the right move, especially if you already use Hermes as a persistent self-hosted assistant. Typing long prompts on a small screen is slow and error-prone