For AI agents: a documentation index is available at /llms.txt. Every page also has a Markdown version at its URL plus .md.
~/what-i-do/ai
AI earns its place when it removes real, repetitive work and you can trust what it does. I build AI features and agent workflows with structure, measurement and humans on the uncertain cases.
The copy-paste between webshop, accounting, spreadsheets and email - done by the systems instead of by people. I find the tasks that eat your week and connect the tools you already have, in code or with low-code tools like Power Automate, Make or n8n when that suits you better.
Product texts, sorting incoming mail, reading invoices and orders into e-conomic, Dinero or your ERP: AI does the routine work, and a person checks only the cases it is unsure of. Tested on your own data before you commit.
An assistant that answers from your own documents, FAQ and product data: draft replies to customer mail for a person to approve, a chat on your website, or search across SharePoint and Drive that shows its sources.
Hands-on use of ChatGPT, Microsoft Copilot and similar tools on your team’s real tasks, plus simple rules for what data may go where. Your staff leave with workflows they use the next morning.
Get real leverage from coding agents without the chaos: skills, hooks and a gate the agent can’t quietly weaken, plus workflows from ticket to reviewable PR.
Schema-enforced output, retries and idempotency, confidence scoring, human-in-the-loop review and an audit trail. Any major LLM provider, behind an abstraction you can swap.
One shared way for every team to use LLMs: a provider-agnostic gateway with cost-aware model routing, usage and spend per team, guardrails, caching and evals, so each new AI feature starts on a paved road.
Let ChatGPT, Claude and other AI agents find and use your product: an MCP server, llms.txt, Markdown versions of your pages, API catalogue and structured data - the same setup this site runs.
Models answer in schema-enforced JSON, so the rest of the system gets typed data instead of guesswork.
Seen in: Illux Product AI → · jev-tools →
Confident results flow through; uncertain ones go to a person. Automation where it’s safe, judgement where it matters.
Seen in: jev-tools → · Illux Product AI →
Fast, cheap models handle the bulk; stronger ones are reserved for what they can’t decide.
Seen in: jev-rerank → · jev-tools →
Evaluation harnesses compare models and prompts against real labels, and audit trails record exactly what ran.
Seen in: jev-eval → · Illux Product AI →
An agent’s risky actions are checked before they run, and when the check is unsure it fails safe: the call is held for a person instead of going through.
Seen in: jev-tools →
Counting, dates, routing and any rule that can be written down stay in plain code. A model only gets the questions that actually need judgement.
Seen in: jev-tools →
Generative-AI product enrichment & room visualization for Shopware 6
A polyglot engineering gate for AI-assisted development
Three TypeScript tools that put a cheap, typed model in front of the expensive one
Keeps a coding agent to the project rules no linter can check
AI job-application platform - from posting to tailored CV
A dependency-free .NET client for Jev
A dependency-free Java client for Jev
Measure Jev on your own labeled data and get the confidence cut-point to gate on
Near-free relevance filtering and reranking for RAG
Fast typed classifiers against Claude, measured on the same requests
A syntax-aware duplicate code finder that hands an agent the helper to extract
What it takes before AI-generated product copy can run in a real webshop - lessons from a solution that has enriched more than 3,000 products.
How to spot the manual work that's worth automating, with five typical examples from small and medium-sized Danish businesses.
Three open-source tools - an agent guardrail, a GitHub issue-triage Action and a bulk classifier - built on one pattern for AI automation that's cheap enough to run on everything.
How a Shopware 6 plugin turned product images into multilingual copy, SEO and categories for Illux - and the guardrails that made it safe to run unattended.
Why I stopped treating CI and "AI coding standards" as two separate things - and what Foundry does instead.