My goal is always to build something that is useful. It is easy to say and hard to do. So the next question is: how? My answer is one word: habits. Pickleball shows this clearly. When I play and try hard to improve, there are lots of things — tricks — to master. Some tricks … Continue reading My Goal Is to Build Something Useful

The real engineering priorities are: A deletion that cannot happen by accident. A result that can be reproduced and replayed. A write operation that cannot be applied twice by accident. An artifact whose complete provenance can be reconstructed. A system that captures its own misses and learns from them — without relying on an agent's … Continue reading Hardening an MCP Server: Five Engineering Priorities, from DeepSeek Super-Agent

The web was originally designed for humans. Websites organize information visually, guide visitors through navigation menus, and expect people to read, click, compare, fill out forms, and contact businesses. AI agents introduce a different kind of visitor. An agent does not necessarily want to browse a website the way a human does. It wants to … Continue reading Building an Agent-Ready Web: From Readiness Audit to a Real Business

Mark Zuckerberg backed this young genius, Alex Wang, with billions of dollars, and in just seven months he delivered a phenomenal agent—Muse—to the market. It’s so sticky that once I used it, I basically couldn’t live without it. I definitely want to dive deep into how it’s built, but unlike Meta’s LLaMA LLM, they haven’t … Continue reading Muse – a Real Agent Built for Consumers

When we use AI agents to perform meaningful work, quality control cannot be an afterthought. We need a systematic way to verify that the output is correct, not merely plausible. At a minimum, I believe there are two essential steps. First, use an independent agent to review the work. Ideally, this should be a different … Continue reading Quality Assurance When Using AI

An LLM generates one token per forward pass. The expensive part is often reading model weights from GPU memory. If one pass serves many requests, the cost of reading those weights is shared, improving throughput. But each request has its own history, which the model needs for attention. KV cache stores the keys and values … Continue reading How vLLM and Deepinfer Serve Many Requests on a GPU

REA (Reverse Engineer Anything) is an AI agent that has recently attracted significant attention. Its idea is simple: give it a binary file, and it can analyze the compiled program and attempt to reconstruct a similar codebase that reproduces its functionality. The impact could be profound. A compiled program is the result of translating source … Continue reading REA: Reverse Engineering Anything

A computer is built from circuits that store bits and move them between memory and processing units. At the hardware level, a memory address identifies a location, and the bits stored there encode data or instructions. But the hardware alone does not explain how a program manages memory, why one data structure performs better than … Continue reading Memory, Data Structures, Caching, and Concurrency

How can these simple components be connected to store data, perform calculations, and execute a program? A MOSFET, or metal-oxide-semiconductor field-effect transistor, is a fundamental building block of modern digital electronics. It acts like an electrically controlled switch: a voltage applied to its gate controls whether current can flow between its source and drain terminals. … Continue reading From MOSFETs to the CPU

The computer brains I am quite familiar with now are Arduino, STM32, and Raspberry Pi. They are all computers, but they sit at very different levels of capability. An Arduino is typically built around an 8-bit microcontroller. An STM32 is generally a 32-bit microcontroller. A Raspberry Pi, on the other hand, uses a 64-bit application … Continue reading The Brains Inside the Computers I Work With