AKAntonios Kioksoglou
Contact
Questions

Questions and answers

Plain answers about AI engineering, Meta Ads and e-commerce, in my own words.

About Antonios

Who is Antonios Kioksoglou?

I’m Antonios Kioksoglou, a software engineer and technical founder in Gothenburg, Sweden. I’m CTO and co-founder of Zalster, a SaaS for optimizing Facebook ads, and I work where AI, e-commerce, marketing technology and product development overlap. The About page has the story, football included, and the candidate summary has the facts in plain text.

What do you do with AI?

I build AI agents and AI-powered products, and I build software with agents. Claude Code and a few narrow sub-agents (a builder, a reviewer and an integrator) write the first draft, while I write the spec and read every diff. The Software page shows the setup, the house rules and an example eval log.

Where are you based, and which languages do you speak?

I’m based in Gothenburg, Sweden. Swedish is my native language, and I work professionally in English and Greek.

AI engineering

What does an AI engineer do?

An AI engineer builds software on top of large language models: agents that use tools, the data and integrations around them, and the tests that show whether the whole thing works. In my own work that means a written spec first, agents with only the access they need, evals that gate changes, and a person reading every diff. The model is one part of the system. Most of the work is the rest of it.

How do I make my site easy for AI answers to read?

Put the text you want quoted in the HTML, not in a script. Give images and embeds their sizes, serve your fonts yourself, give every page its own title and description, and keep structured data to what the page says. Google’s own guide says there are no additional requirements for its AI features and that Google Search does not use llms.txt. The article has the numbers from this site and a ten-minute test.

How do you stop an AI agent from breaking things?

With four rules: spec first, every diff read, read-only by default, and tests and evals that gate every change. Agents never get production. Writes to live data go through a reviewed path, and anything that touches money, permissions or deletion is read twice. The reasoning is in How I work on the Software page.

What is an MCP server, and why would an agent need one?

MCP, the Model Context Protocol, is a standard way to give an agent tools and data. An MCP server exposes a narrow set of actions, such as reading an account or a support log, so the agent can work with a store, a CRM or a database without being handed the keys to everything. The Software page has an example, and this article shows a read-only setup for GA4, Google Ads and BigQuery.

What are Claude Code mods?

Mods are small TypeScript functions, announced on October 1, 2026, that change how Claude Code works: they can rewrite a prompt, block or change a tool call, draw their own panel and replace built-in features. They ship inside plugins and run with the same access to your machine as Claude Code, so install them only from people you trust. The article explains what they are and builds a small one that shows an article’s reading time while Claude edits it.

How are Claude Code mods different from hooks?

Anthropic’s announcement says hooks cannot rewrite an event, draw new UI or replace a feature, and mods can. The getting-started guide describes mods as hooks that ship inside plugins and see every event in your session. The article has a short comparison table.

How do I build a Claude Code mod?

Describe the mod to Claude Code and allow hot reloading when it asks, or write it yourself: a plugin folder with a manifest, a hooks.json that names one module, and a module that exports register. Check it with claude plugin validate and claude plugin test. The article builds a small one, with its code and the checks I ran.

Are Claude Code mods safe to install?

Not automatically. A mod runs with the same access to your machine as Claude Code itself and is not sandboxed, so install mods only from sources you trust and read the source first. On Team and Enterprise plans, a built-in mod called sec-default loads first and stops installed mods from risky behavior such as overriding permission deny rules. The article lists what to check.

What is the Cloudflare Web Search API?

It is a beta service, announced on October 2, 2026, that lets an AI agent or application search the Internet and use live results instead of relying on a model’s training cutoff. Requests go through Cloudflare’s AI Gateway to one of three search providers: Ceramic.ai, Exa or Linkup. The article explains what it means for you, how to call it and what you can do with it.

How do I use the Cloudflare Web Search API?

Use an AI Gateway that has credits or a provider key, then send a POST request to your account’s /ai/websearch/ endpoint with a Cloudflare API token, or call env.AI.websearch() from a Worker. A query can be up to 1,024 characters and a request returns up to 10 results, each with a URL, a title and a description. The article has the steps, a table of the fields and a small tested script.

How much does the Cloudflare Web Search API cost?

Cloudflare says searches are billed to your AI Gateway credits at each provider’s list API price, with no additional markup, or you can bring your own provider key and the provider bills you directly. The two Cloudflare pages I read give no prices, so check each provider’s list price before running many searches. The article shows how to keep a test run small.

How can I check whether my site shows up in AI search results?

Ask the questions your customers ask through each search provider and record where your domain appears in the results. The Cloudflare Web Search API makes that one request per question and provider, but its three providers are not Google or an AI assistant, so treat the answer as one view, not the whole picture. The article has a small script with tests.

What is new in Claude Haiku 5.5?

Claude Haiku 5.5, released on October 7, 2026, lists at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, down from $1 and $5 for Haiku 4.5. Anthropic calls it its fastest model, and it is the first Haiku with an adjustable effort setting, a 1 million token context window and up to 128,000 output tokens. The article has the benchmarks and what to watch for.

Which Claude model should I use: Fable, Opus, Sonnet or Haiku?

Start with Claude Opus 5.5 for most workloads, as Anthropic’s docs advise. Use Sonnet 5.5 for faster everyday coding and agent work, Haiku 5.5 for high-volume, latency-sensitive jobs such as classification, extraction and routing, and Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort falls short. The article has a table with examples from store and marketing work.

Can one Claude model orchestrate the others?

Yes, but you assign the roles. In Claude Code you can use opusplan, which plans with Opus and executes with Sonnet, set a model on a subagent, add a stronger advisor model, or run a dynamic workflow. On the Claude Platform, Managed Agents has a coordinator that delegates to worker agents, each with its own model. The article says when Anthropic’s own measurements found it pays.

When is Claude Haiku 5.5 the wrong choice?

For complex agentic coding. Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choices there, and on its Terminal-Bench 4.0 benchmark Haiku 5.5 scores 39.2 percent against 70.6 for Sonnet 5.5. Haiku 5.5 suits narrower work such as compaction, summarization and subagents. The article lists the other limits.

What is ChatGPT Intelligent UI?

ChatGPT Intelligent UI is a capability of GPT-6 in ChatGPT, announced by OpenAI on October 7, 2026, that lets it answer with an interface instead of only text: graphics, tappable buttons, forms, charts and small tools you use inside the conversation. The model chooses the format from your question, and can still answer in plain text when that is most useful. The article explains it, and a page of mock-ups shows the idea.

Who gets ChatGPT Intelligent UI, and when?

OpenAI says GPT-6 with Intelligent UI started rolling out globally on October 7, 2026, to ChatGPT Plus, Pro, Business and Enterprise in the Chat tab, and expands to Free and Go the next day. Enterprise availability depends on workplace admin settings. GPT-6 Sol powers Plus, Pro, Business and Enterprise, GPT-6 Luna powers Free and Go, and the models behind Work and Codex are not changing. The article has the facts in one list.

How does ChatGPT Intelligent UI work?

OpenAI describes a library of native, streamable components and a compiler that processes the interface as the model generates it, so an answer appears piece by piece. GPT-6 was trained to decide what to say, how to lay it out and what to make interactive, and the interfaces were evaluated for clarity, usefulness and completeness. The post does not list the components or publish the evaluation. The article has a figure of the flow.

Is ChatGPT Intelligent UI available in the OpenAI API?

Not that I can find. OpenAI’s October 7 post is about the Chat tab in ChatGPT and does not mention the API, and its October 2 model guide for the GPT-6 family, written for API and Codex users, does not mention Intelligent UI either. Check OpenAI’s documentation for your date before you plan around it. The article lists what the announcement does not say.

What is Mistral Timestral?

Mistral Timestral is a time-series foundation model in four sizes, released on October 15, 2026. It is pre-trained on many different series, so it forecasts a new series without being trained on it first, and it returns a range of 21 quantiles for each future step instead of one number. It can use related variables and covariates, such as a holiday calendar. It is not a language model. The article covers what it does and where it could help e-commerce, tech and marketing.

Is Mistral Timestral free to use?

Partly. The code is Apache 2.0, but the model weights come under a modified MIT license with a revenue limit: if your company’s or your employer’s global consolidated monthly revenue for the preceding month exceeded $20 million, you have no rights under it unless Mistral grants you a commercial license, which it may refuse, or you use its hosted services. Below that limit the license allows use, modification and sale with the notice kept. I am not a lawyer, so read the license file. The article has the details.

Does Mistral Timestral need a GPU?

No. Mistral says Timestral runs on a CPU. In its own test, on a server with 8 threads, a one-day-ahead forecast from three weeks of 15-minute data took 10 ms for the 5M model and 140 ms for the 209M one. The four sizes are 5M, 28M, 123M and 209M parameters, with weights from 21 MB to 837 MB. The article has a table of the sizes.

Can Timestral forecast e-commerce demand?

Possibly, but Mistral does not say. Its post’s worked examples are industrial and finance. What fits is its support for known-future covariates, such as a promotion calendar, and a range for each step. On the four sales configurations in GIFT-Eval, which give no promotion calendars, I computed a MASE about 32 percent lower than seasonal naive, a simple baseline, so backtest it against what you use today. The article shows how.

What is Mistral Large 4 (Le Chonk)?

Mistral Large 4, nicknamed Le Chonk, is a mixture-of-experts model with 1.05 trillion parameters, about 52 billion of them active per token (49 billion without embeddings), that accepts text and images and answers in text, with a 1M-token context and three reasoning levels. Mistral ran a preview on its API from October 6, 2026 and released the weights on October 29. The article covers the license, price, benchmarks and where it could help.

Is Mistral Large 4 open source, and can I use it commercially?

Mistral calls it open-weight. The weights come under the Mistral Ecosystem Open-Weight License, a modified MIT: free to use, modify, fine-tune and sell with the notice kept, except that a company that runs a model-as-a-service business, giving third parties access for inference or post-training with meaningful control over inputs, parameters or training data, must take a commercial license from Mistral. Internal use, features that only embed the model and use through Mistral’s own products are outside that rule. I am not a lawyer, so read the license. The article has a chart.

How much does Mistral Large 4 cost, and what hardware does it need?

Mistral’s API charges $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Self-hosting is heavy: the FP8 weights are 1.06 TB and Mistral’s recipes use eight-way tensor parallelism, while a 610 GB NVFP4 checkpoint targets Blackwell GPUs. Mistral states no minimum hardware. The article has the arithmetic.

How good is Mistral Large 4 compared with other open models?

Mistral says it is the top open-weight model for coding, and its own charts show it ahead by 0.4 points on FrontierCode 2.0 and 1 point on Terminal-Bench 4.0. Artificial Analysis’s own runs of those tests put it behind DeepSeek V4.1 Flash 1017 and GLM-5.3, and its Intelligence Index score is 26 against 32 for both. Mistral’s model card calls it the highest-scoring Western open-weight model on that index and says the strongest from China score higher. The article has the table.

What is the Gemini Decisions API?

The Gemini Decisions API is a Google endpoint, introduced on November 11, 2026, that answers closed questions instead of writing text. You send one input (text, images, audio, video or PDFs) and up to 1,000 yes/no, multiple-choice or rubric-score questions, and its model, Gemini Decision 1, returns a probability for every option and a confidence, with zero output tokens. Google quotes 150 to 250 milliseconds end to end for inputs up to 16K tokens. The article covers the price, the benchmarks and how to pair it with an LLM.

How is the Gemini Decisions API different from a normal LLM?

An LLM writes; Gemini Decision 1 chooses. It scores the options you supply in a single pass, so your code gets a probability to branch on instead of text to parse, and it does not write summaries, call tools or reason over many steps. Google’s docs send those jobs to its Interactions API. In Artificial Analysis’s own run, two language models in versions it labels Low score 1 to 3 points higher on its decision index but take 14 to 21 times longer per task, and the Max versions score 10 points higher at 54 and 153 times longer. The article shows how to use both.

How much does the Gemini Decisions API cost?

Gemini Decision 1 costs $0.10 per million input tokens with nothing billed for output, and the shared input of a request is billed once however many questions you ask. The free tier in Google AI Studio lets Google use your data to improve its products; the paid tier does not. There is no Batch API or caching, and Tier 1 allows 1,500 requests a minute and 50,000 a day. The article has a made-up example.

How can I use the Gemini Decisions API with an LLM?

Let it make the quick, repeated decisions around the LLM. In my view that means a front desk that routes each message to a template, a small model or a big one; an editor that checks drafts against your rules; a checklist that an LLM writes from your guide and the decision model runs on everything; a labeler that tags thousands of items before an LLM summarizes them; and brakes for an agent. Google says Gemini Notebook routes queries this way. The article has a figure and a code sketch I have not run.

Marketing analytics and Meta Ads

How can AI help with marketing analytics?

Mostly by making small decisions cheap: sorting leads, search terms, reviews and campaign names into clear categories that people can act on. I wrote up seven such jobs and the checks to run first. For the numbers themselves, the sample reporting workspace shows what I look at first in an account: revenue, spend, MER, new-customer cost and what to fix first.

What are Claude Dashboards?

Claude Dashboards is a beta feature, announced on October 8, 2026, that turns a plain-language question about your data into a dashboard. It connects to a data warehouse such as BigQuery or Snowflake, or to an app you have connected, such as Salesforce, and each chart shows its SQL query and when its data was last refreshed. The article covers what it could change for a marketing team.

What is Claude Motion?

Claude Motion is a beta feature, announced on October 8, 2026, that turns your content into a short animation, such as a results explainer or a product walkthrough. It is built as code that animates your own text, charts, shapes and images, so you can change a word, a number or a timing, and it exports as an MP4. It does not use a video generation model. The article has the marketing uses.

Who gets Claude Dashboards and Claude Motion?

Both are in beta. Claude Dashboards is available on Pro, Max, Team and Enterprise plans, and Claude Motion on Team and Enterprise plans. On Enterprise plans an owner turns them on in Organization settings, and Motion is on by default on Team. Both count toward your plan’s usage limits. The article has a table of the two.

Can I share a Claude dashboard with a client outside my company?

Maybe. Anthropic’s announcement says dashboards can be shared outside your organization, or with anyone who has the link, if your admin allows it. The Help Center’s Share artifacts page and admin guide say artifacts that connect to your apps cannot use anyone-with-the-link sharing, and that people need a Claude account to open one. A shared artifact also uses the viewer’s connections, not yours. No page I read says which rule governs a dashboard built on a warehouse, so test the link before you promise it. The article lists what to check.

What does a marketing analyst do?

A marketing analyst turns spend, traffic and order data into decisions: which channels earn their budget, which campaigns to fund or cut, and what the numbers cannot tell you yet. For an e-commerce brand that means checking the tracking first, then revenue, ad spend, MER, new-customer cost and a short list of what to fix. I do this work for e-commerce brands, and the sample reporting workspace shows how I look at an account.

What is MER, and why track it next to ROAS?

MER, the marketing efficiency ratio, is total revenue divided by total ad spend across every channel. ROAS is reported by each platform and counts only the orders that platform claims, so the platforms can add up to more revenue than the store took. MER uses the store’s own revenue, which makes it blunt and hard to attribute away. A MER of 4 means the store took in four euros for each euro spent on ads.

How do you decide where an ad budget should go?

I start from the store’s own orders, not the platform reports, and do the data analysis there: revenue, spend and MER by week, then the cost of a new customer and the margin and returns behind it. After that comes marginal return, meaning what the next euro buys in each channel, because the average hides it. At Zalster a Bayesian algorithm and regression analysis do this for Facebook ads, and on a small budget I judge changes over weeks, not days.

Which analytics and reporting tools do you use?

GA4 for the event model and e-commerce reporting, and Piwik PRO where a client wants analytics, tag management and consent in one suite with a choice of where the data is stored. BigQuery holds the raw data, and Looker Studio, or Looker Studio Pro when a team needs shared workspaces and scheduled delivery, turns it into dashboards. The Analytics tab on the Marketing page shows how each is used and checked.

What happens to analytics when a visitor declines cookies?

It depends on the tool and the setup. With GA4’s basic consent mode nothing is sent. With advanced mode GA4 receives cookieless pings and can model the missing visits, but only above Google’s thresholds: 1,000 events a day with consent denied and 1,000 consenting users a day. Piwik PRO can record anonymous visits instead, with the IP address masked, the country only, and sessions tied together by a hash or a 30-minute cookie. Whether that is allowed is a question for a lawyer. This article compares the two.

What is server-side tracking, and does a store need it?

Server-side tracking sends events such as purchases from your own server to the ad platform, not only from the visitor’s browser. Browsers block scripts, cookies expire and consent is required, so browser-only tracking misses orders. The usual fix is the browser pixel plus Meta’s Conversions API, with each order sent once and matched by an event ID. The checklist for a store is in this article.

How many campaigns and ad sets should a small store run on Meta?

Start with one campaign and one ad set, optimized for purchases. Meta needs roughly 50 optimized events a week per ad set to leave the learning phase. On the conservative assumption that a sale costs about what the product sells for, a store with a €60 average order needs about €430 a day to get there. Spread thinner than that, one ad set is better than five. The arithmetic is in my article on Jon Loomer’s Master Brief.

How do I know whether Meta’s reported results are real?

Compare them with the store’s own orders every week, expect the two to differ, and look for a boring explanation before suspecting Meta: an event set up on the wrong page, a deduplication problem, or one person buying several times. Treat purchases counted from a view as weaker evidence than purchases counted from a click, and keep one number that no platform can attribute away: total revenue divided by total ad spend. The full method is in my article on reconciling Meta, GA4 and the store.

What should I check first when Meta results drop?

Ask three questions in order. Why are people not clicking? That is the creative and the offer. If they click, why do they not buy? Look at the store: product page, cart, checkout, shipping costs and stock. If they buy, why does Meta not know? Check the events, deduplication and consent. Most fixes turn out to be in the store, the tracking or the creative, not in a campaign setting. The full walk-through is in the same article.

How do you measure ChatGPT Ads?

Create a data source in Ads Manager and send conversion events with the OpenAI Pixel, the Conversions API or both, using the same event ID for the same conversion. Then choose reporting windows: 7, 14 or 30 days after a click, and 0 or 1 day after a view. OpenAI’s October 5, 2026 post adds partner integrations for attribution and incrementality. The article has the steps and a four-week test design for a store.

Is ChatGPT Ads available in Sweden?

Yes. OpenAI’s Help Center lists Sweden as available for self-service in Ads Manager, and its August 18, 2026 announcement named Sweden among 31 European markets. The legal entity that advertises and is billed must be based in a listed country, and ads appear only to people on the Free and Go plans. The article says what that means for measurement.

How should I read the results OpenAI published for ChatGPT Ads?

As cases reported by partners, not as a benchmark. The five results in OpenAI’s October 5, 2026 post come from DV Rockerbox, WorkMagic, INCRMNTAL, Triple Whale and Fospha, each measuring something different, and the post gives no sample sizes, dates or spend. OpenAI’s own FAQ says ChatGPT Ads has no performance benchmarks yet. The article has a table of what each number does and does not show.

Do ChatGPT Ads attribution windows change how a campaign optimizes?

No. The 7, 14 and 30-day click windows and the one-day view window change reporting only. What a campaign learns from depends on its billing: with click billing it optimizes toward conversions after a click, and with impression billing toward a broader set that includes conversions after a view, according to OpenAI’s Help Center. The article explains the difference.

E-commerce

Which e-commerce platforms do you work with?

Any of them: hosted platforms such as Shopify and BigCommerce, open-source ones such as WooCommerce and Magento, and headless builds, plus the payments, ERP, shipping, feeds, search and tracking around the store. The E-commerce page lists the whole stack and lets you search for the tool you use.

Can I see traffic from ChatGPT and other AI assistants in GA4?

Partly. Since May, GA4 has a default AI Assistant channel for visits from sources like ChatGPT, Gemini and Copilot. It reads the referrer, so a visit without one is filed as Direct, and clicks from Google’s AI Overviews and AI Mode stay in Organic Search. Treat the number as a floor. A custom channel group catches other assistants and older data. The article on product feeds and GA4 has the setup and a feed checklist.

Where do online stores usually break?

At the checkout, in the catalog data and in the integrations, and usually on a Friday. The common ones are payments that double-charge or drop orders, feeds with bad product data, and syncs between the store, the ERP and the warehouse that have no retries and no readable logs.

Working together

How do I get in touch?

The candidate summary has my email address and LinkedIn. Tell me what the project is and I’ll say honestly whether I’m the right person, including when the answer is no. The CV is there too, as a PDF.

Are the numbers on this site real?

The numbers on the Marketing page are made up, and the terminal sessions on the Software page are scripted simulations. Neither is a client result. The methods are the ones I use on real accounts, and the candidate summary names the companies I’ve worked with, by name only.