AI EngineerBern, Switzerland
I test AI end to end: local LLMs, my Personal Agent, and AI in apps and websites that are live. Every week I test one AI idea in public and show what worked and what didn't.
What comes after the demo works: cost, reliability, privacy and store review.
Two store apps run LLM features today, Linki with 135K+ installs. Linki's AI link analysis costs about $0.002 per link, with per-user limits, capped refunds on failed calls and a monthly spend brake. Rumbelly's photo scan costs about $0.0024 per scan, behind App Check and a server-side scan limit.
Production incidents traced from the logs to the root cause and fixed: a silent rate limit, a billing stop at the model provider and a slow input sanitizer caught in security review.
App Store rejections recovered, over subscription terms and permission prompts. Privacy labels built from what each app actually sends. Private user data kept away from the model.
Built with Claude Code agents, MCP servers and local models. Research, design, code, tests and release all run through agents.
Linki and Rumbelly are my own products. I build and run both.
Live products first, then what is in build. Each page links to the website or the store listings.
Bookmark manager for Android and iOS with AI summaries, automatic tags and voice search. 135K+ installs.
Gut-health tracker with an AI meal and menu scan. Live on the App Store and Google Play since August 2026.
A new app in build. Its name, icon and details stay under wraps until it's in the stores.
A new app in build. Its name, icon and details stay under wraps until it's in the stores.
A weekly show about practical AI. I try one idea on a real task, then decide: is it worth shipping, or should you skip it? Notes from each experiment go out on LinkedIn and X too.
First episode coming soon on YouTube.
Give AI the goals, decisions and limits of a project before it acts, so you explain less and correct less.
The main thread
Personal apps and chats for home, work and career, connected to that same context.
For everyone
Project structure, guardrails, security and token budgets that make agent work useful.
For people who code a little
In the works
A small folder of Markdown context files and agent instructions, so ChatGPT, Gemini or Claude know a project's goals, decisions and limits before they start. I'm building it from the system I use every day.
I'm an AI Engineer (Mobile) in Bern, at one of the biggest companies in Switzerland. I've written native Android since 2009 and, since 2023, Kotlin Multiplatform apps that run on Android and iOS from one codebase. That's where my AI work gets tested: real users, real costs and store review.
I've been building with AI since 2021, starting with GPT-3 and Copilot, and my whole workflow has been AI-first since 2024. Today I build with Claude Code agents, MCP servers, local LLMs and my own Personal Agent, from research and design to code, tests and release. What ships is checked: measured quality, predictable cost, privacy by design.
My applied AI master's at TheBigSchool covers LLMs and agents, MCP, AI law and AI for marketing.
AI first, with the mobile depth it runs on and the web that ties it together.
Daily-driven multi-agent stack, in production.
Native + cross-platform shipping for Android and iOS.
Modern web for landing, dashboards and APIs.
A practitioner programme on applied AI: LLMs and agents, MCP, vibe coding, AI law (the EU AI Act) and AI for marketing.
Erik is an AI Engineer in Bern, Switzerland. He builds and tests AI end to end, from local LLMs and his Personal Agent to AI features in live apps and websites, and shares practical AI experiments in public.
Linki, an AI bookmark manager for Android and iOS with 135K+ installs, and Rumbelly, a gut-health tracker with an AI meal and menu scan. Both are his own products; he builds and runs them.
About $0.002 for one Linki AI link analysis and about $0.0024 for one Rumbelly photo scan. Per-user limits, a server-side scan limit and a monthly spend brake keep the bill predictable.
A weekly show about practical AI. Each episode tests one AI idea on a real task and ends with a verdict: ship it or skip it. The first episode is coming soon on YouTube.
By email at meleaiautomations@gmail.com or on LinkedIn. He usually replies within 24 hours, Bern time.
