Technical notes from our Pilot Builds.

Reference architectures, technology selection criteria, and rollout roadmaps — written for people who'll actually read the code and run the system, not company news.

DevelopmentOpen SourceWindowsDeveloper ToolsAI

Action Wheel: a radial menu for Windows — one middle click puts your everyday actions right under the cursor

Action Wheel is KonexForge's open-source (MIT) app for Windows 10 and 11: middle-click anywhere and a ring of 9 buttons appears around the cursor to send a shortcut, launch an app, run a Windows function, or open an AI tool such as Claude Code or Codex. This post covers how it works, how we designed the AI features to be safe, privacy, and how to install it in a few minutes.

28/10/202610 min read

DevelopmentPHPNode.jsPatch managementAI

PHP2Shell and Node2Shell: flaws that are neither new nor old, and why the patch is only the start

On 21 October 2026, PHP and Node.js both shipped emergency fixes for remote code execution flaws in their core — PHP2Shell (CVE-2026-109618, CVE-2026-109619) and Node2Shell (CVE-2026-111497). Both were found with the help of AI, in code that had been sitting there for years. This post sums up what happened, why updating one server does not mean you are patched, and a checklist for businesses running PHP or Node.js.

24/10/202614 min read

AI & MLLivestreamDigital humansAIE-commerce

AI virtual KOLs livestreaming 24/7: how far China has gone, and what bringing it to Vietnam takes

In China, AI livestream hosts already sell real goods: Luo Yonghao's AI avatar brought in 55 million yuan of orders in a single session, and more than 70,000 sellers on JD use virtual hosts. But the rules have tightened too: AI content must be labelled, a real person must supervise, and the operator answers for every word the virtual KOL says. This article sums up what China has done, Vietnam's legal framework in 2026, and how to run a virtual KOL within the law — a service KonexForge is ready to deploy for businesses in Vietnam.

18/10/202613 min read

Optimization LoopWebsite speedOptimization LoopMeasurement

Fast on your device, slow on your customer's phone: measure with real-user numbers before you fix anything

A business owner opens their website on a new phone over office Wi-Fi and sees the page in a second and a half. A customer on a mid-range phone over 4G can wait twice as long. Google judges speed by the 75th percentile of real users, not by what the site owner sees. This article explains the three Core Web Vitals in plain language, why a Lighthouse score is not one of them, three calculations you can do yourself, and six steps to measure, fix and measure again.

18/10/202615 min read

DevelopmentVietnameseAI costsUnicodeSoftware development

How many tokens does Vietnamese cost? Measured on six tokenizers — plus a formatting issue that multiplies the AI bill by two and a half

"Using AI in Vietnamese costs more than in English" is true, but how much more depends on the model. We measured eight bilingual business texts on six public tokenizers: with newer tokenizers, Vietnamese costs about 1.2–1.4 times as many tokens; with older ones, about 2.1–2.3 times. Typing without diacritics saves nothing on newer tokenizers — it costs more. And if Vietnamese text is stored in decomposed Unicode form (NFD), the token count can rise nearly 2.5 times, fixable with one line of code.

17/10/202611 min read

Data AnalyticsBenchmarkMeasurementStatisticsAI

Benchmark scores are not your scores: build your own evaluation set to choose an AI model — and how many questions it needs

Every week a new AI model claims the top of some leaderboard. But leaderboards measure other people's work, and the two biggest ones right now do not even agree with each other. The only way to know which model is good for your work is to measure it yourself. This article shows how to build an evaluation set from real data, and uses simulation to answer a question few people calculate: with two models 6 percentage points apart, 30 questions pick the better model only about 72% of the time — you need around 200 to be nearly certain.

16/10/202611 min read

AI & MLOpen-weightSelf-hosted AIGPUAI costs

Open models in October 2026: free to download, but is running them yourself cheaper than an API?

From August to mid-October, Alibaba, DeepSeek, Xiaomi and Meta each released AI models anyone can download for free, some with a trillion parameters. Many businesses read the news and think about running one themselves to stop paying for an API and keep data in-house. This article does the maths: a leading open model needs about $25,000 a month in GPU server rental, while the same model through an API costs a few tens of dollars for a mid-sized business's volume. There are good reasons to self-host — but saving money is rarely one of them.

15/10/202611 min read

AI & MLClaude HaikuGPT-6 LunaGemini FlashAI costs

Small AI models in October 2026: Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash — dozens of times cheaper, and three traps in the price list

The newest small models from Anthropic, OpenAI and Google reach roughly 70–80% of a flagship's score on an independent index, at a cost per task that is a few dozen times lower. For most of a business's repetitive work — sorting messages, extracting details, drafting standard replies — that is enough. But their price lists hide three traps: a price cliff at 100,000 tokens, a promotional price about to end, and a reasoning setting that can change the cost fifteenfold on the same model. This article works through each one and shows how to test whether a small model is enough for your work.

14/10/202612 min read

AI & MLGemini 4GPT-6ClaudeAI costs

Gemini 4 Argon, GPT-6.1 Sol, Claude Sonnet 5.5: the same $2/$10 list price, nearly seven times apart per task

In the last two weeks of September and the first days of October, three major labs each shipped a model with an identical list price: $2 per million input tokens and $10 per million output tokens. Yet on the same independent benchmark, the real cost of finishing one task differs by nearly seven times, the two leading leaderboards name different winners, and one of the three prices doubles after 30 November. This article reads the numbers closely and shows how to work out your own cost per task before choosing.

13/10/202611 min read

DevelopmentAISecuritySoftware development

Vibe coding: building an app in a weekend is real — and so are the five problems that come with it

Vibe coding — describing what you want, letting AI write all the code and not reading it — lets people who cannot program build working apps in hours. But Veracode's research found 45% of AI-written code samples failed security testing, nearly one in five libraries AI suggests does not exist, and one common misconfiguration exposed the data of 170 apps. This piece separates when vibe coding is worth using from when it is not, with six self-checks before an AI-built app goes live.

10/10/202615 min read

Data AnalyticsData analyticsExcelRetail

Reorder on average sales and you run out every other cycle — working out safety stock in Excel

The most common way small shops and distributors reorder is average daily sales times the days the delivery takes. We simulated 200,000 replenishment cycles: that rule runs out of stock 50% of the time. This piece gives a safety-stock formula that fits in six Excel cells, shows why a supplier who delivers two days either side can double the buffer you need, and sets out when the formula does not apply.

04/10/202612 min read

AI & MLClaudeAI costsAI safety

GPT-6 versus Claude Opus 5.5: ten days after the call to slow down, three new models in one afternoon

On 12 September Anthropic's CEO called on the industry to slow down and OpenAI's CEO agreed within hours. On 22 September Anthropic released Claude Opus 5.5, and minutes later OpenAI released GPT-6 Sol and Luna. This piece compares the GPT-6 family with Opus 5.5 using numbers labelled by who measured them, shows why the slowdown call cannot be broken — it contains no number — and sets out what a business should do when models change by the month.

27/09/202618 min read