Kenji Tanaka is a RECATOOLS editorial persona focused on developer tools, cloud platforms, DevOps, CI/CD, software supply chains, and infrastructure trends. Articles under this byline help technical readers understand how tooling changes affect software delivery, security, cost, and reliability.

About this byline

Kenji Tanaka is a RECATOOLS editorial persona for developer tools, cloud, DevOps, and software supply-chain coverage. Articles are produced and reviewed under RECATOOLS editorial supervision.

Editorial responsibility for this byline rests with Jeffrey Tan (Tan Seng Fei), RECASYS, Singapore. Our Editorial Policy sets out how sources are reviewed, how AI-assisted workflows are used, and how editorial desks and personas are defined.

208
Articles
Developer Tools
Primary beat
Apr 2026
Writing since
~1664 min
Total reading
All 208 Developer Tools 73 AI & ML 53 Cybersecurity 42 Cloud & Infra 34 Cyber Threat Intel 3 Consumer Tech 2 Science 1

Articles · AI & ML Showing 31–40 of 53

A row of identical unlabelled cardboard boxes on a dim warehouse shelf, lit by a single work lamp from the left
AI & ML

How to Choose an AI Tool Without Believing the Marketing

Five checks worth running before you adopt anything, and where each answer actually lives — because none of them are on the landing page and two of them are not in our own directory either. Worked through real cases: the product rated 4.7 and 1.6 by two different review sites, the platform that cut included credits by 62% while raising the price, and the free tier worth $140 a month that is not a trap at all.

1 Aug 2026 · 8 min read
An old hardback book lying open, its thin pages caught mid-turn in warm light.
AI & ML

Why It Confidently Makes Things Up

We asked two AI models for five peer-reviewed studies on an obscure topic, with DOIs. One produced five — authors, journals, years, and a summary of what each paper found. None of the five exists; every DOI is dead. But every invented DOI used a real prefix belonging to the correct Taiwanese registrar, with only the article identifier made up. That is the whole phenomenon in one detail: the shape is right, only the existence is wrong. Why this is the machinery working normally rather than malfunctioning, why obscure and recent and precise questions are worst, why the other model declined, and the four checks that actually catch it — including the one we nearly got wrong ourselves.

29 Jul 2026 · 9 min read
The wooden drawers of an aged library card catalogue with brass fittings.
AI & ML

MCP, RAG, Agents, Skills: The New AI Vocabulary in Plain English

Somebody says the agent will use MCP to hit the connector, then RAG over the docs, and the skill handles the rest. Everyone nods; most are guessing. Here is every term — LLM, AGI, RAG, tools, MCP, connectors, agents, workflows, computer use, CLI, skills and memory — explained with no prior knowledge, using a single idea: an LLM alone can only write, and everything else is a way of giving it more reach. Ends with the part that matters commercially — why a system that cannot tell your instructions apart from text it merely read is a security problem by design, illustrated by the April 2026 attack that hijacked three coding agents through a pull request title.

26 Jul 2026 · 13 min read
Rows of set metal letterpress type, the letters mirror-reversed as they are on real printing sorts.
AI & ML

Tokens and Context Windows, Explained From Zero

You are billed in a unit nobody explained. A token is a countable stand-in for a chunk of text — the word is borrowed from compilers, where tokenisation has meant this since the 1960s — and chunks are decided by how common they are, not by meaning. That is why London is one token and Johor Bahru is four, why ten characters of Japanese politeness cost less than two English words, and why 謝謝你 costs more than 谢谢你. There is no standard across models: the same Thai greeting more than halved between two OpenAI generations, and only OpenAI publishes its tokeniser at all. Plus where the price actually comes from — you are renting GPU time, cached input costs up to 120× less than fresh, and output is dearer because generation stalls on memory bandwidth — what a context window really holds, and why every turn of a chat quietly re-sends the whole conversation.

26 Jul 2026 · 22 min read
A printed invoice on a desk under lamplight, its line items running off the bottom of the page.
AI & ML

What an LLM Actually Costs You

The rate per million tokens is the one number you cannot change, and the smallest part of the bill. Measured on 102 parallel strings of this site's own copy in eight languages: the non-English token tax roughly halved between tokeniser generations — Thai fell from 4.19× to 1.77× — and closed for nobody, while Traditional Chinese costs about 21% more than Simplified for identical content. Plus a post-mortem on our own token counter, which under-estimated CJK by up to 2.81× until today, and the three costs no rate card mentions: conversation re-sent every turn, re-embedding, and paying for failed attempts.

26 Jul 2026 · 10 min read
A wall of stacked shipping containers seen flat-on, forming a grid of doors in pale green and grey
AI & ML

Serve an LLM with vLLM on CPU in Docker: Learn the Production Engine Without a GPU (2026)

We ran vLLM v0.25.1's official CPU image in Docker on Apple Silicon, dodged a 10 GB CUDA decoy, survived three traps, and served Qwen3-0.6B over a real OpenAI-compatible API — at a derived 12.5 tok/s.

22 Jul 2026 · 23 min read
Watercolour paints and brushes arranged on a surface.
AI & ML

Run ComfyUI on CPU in Docker: Node-Graph Image Generation Without a GPU (2026)

We pinned ComfyUI v0.28.0 in a python:3.13-slim container, hit a torchaudio ABI trap and an 8 GB OOM kill, and still rendered a real image in 51 s on CPU.

22 Jul 2026 · 16 min read
Multiple railway tracks with signals and buildings in an urban setting during the day.
AI & ML

Self-Host LiteLLM with Docker: One OpenAI-Compatible Endpoint for Every Model (2026)

We ran LiteLLM v1.93.0 against a local Ollama model in Docker: pinned tags, a 4-line config, token metering, and the wrong answer that proves a point.

22 Jul 2026 · 14 min read
A black and white close-up of a studio microphone.
AI & ML

How to Transcribe Speech Offline with whisper.cpp in Docker (2026)

We ran whisper.cpp v1.9.1 in Docker: an amd64-only tag trap, a SIGILL crash, a gcc fp16 build fix — then 7.3 s of speech transcribed in 1.25 s on CPU.

22 Jul 2026 · 15 min read
A software developer working at a computer.
AI & ML

How to Run Aider With a Local Ollama Model in Docker (2026)

We ran aider v0.86.2 in Docker against a 986 MB local Qwen model. Same prompt twice: one hallucinated diff, one clean auto-commit. Context decided it.

21 Jul 2026 · 17 min read
Editorial Policy →