An angle, a velocity, and the wind.
My first lines of code ran on an Amiga 500. Not long after, on a beige PC, a hidden QBasic game called GORILLAS.BAS became my first real playground: set the angle, set the velocity, read the wind, and let the banana fly. The game shipped as plain BASIC source, so my first hack followed quickly: wilder wind, bigger explosions, a faster banana.
Three and a half decades later the instinct hasn't changed. Give a system the right angle, account for the wind, let it go. Then watch where it lands and adjust. And when the wind itself is the problem, I open the source and change it.
Hacking since school, engineering professionally since 2001, AI-productive since 2017. I build multi-agent systems inside my own companies, write about them on LinkedIn, and founded the AI Meetup Frankfurt.
TAP A GORILLA · SET ANGLE AND VELOCITY · MIND THE WIND
Good players, every one. I don't bet on a single horse: I route between them and keep the local one for when data can't leave the building. Putting them to work is the easy part. Reading the wind is still my job.
Seven levels from my own companies and private labs, oldest first. My own data, my own responsibility, my own oversight.
My 2015 startup experiment: a website that sold craft beer as a subscription. I ran it end to end myself, marketing and fulfilment both, minimum product first and real customers early. The method got written up in t3n as "Lean Launch".
READ IN T3N →The savedroid app launched in 2016 and passed 100,000 downloads, connected to customers' bank accounts through a PSD2-compliant banking integration. In 2017 a production AI liquidity forecast went live on top of it: classic machine learning on account movements, running in a consumer app years before large language models arrived.
As savedroid's CTO I owned the blockchain handling behind the token sale. The interactive ICO website I built held five-figure concurrent viewers at peak load.
Running since 2018, with Home Assistant added in 2020 as just one layer of many: around 500 devices, 2,000 entities and 150 hand-written automations, with Gemini as the model layer for the harder decisions. Lived daily operation, not a whiteboard sketch.
I have finished multiple Ironman races, and I coach myself with a system I keep rebuilding. Four generations, from a plain ChatGPT prompt to a Claude-based multi-agent setup, with specialist agents for run and workout analysis, exercise-video review, shoe management and muscle mapping.
The full MiCAR application, including the DORA operational-resilience requirements running alongside it, was moved into a knowledge base and checked by AI agents for consistency, gaps and regulatory risk.
An eight-year dispute with tens of millions of euros at stake, spanning hundreds of briefs, exhibits and letters, worked through with a multi-agent system that surfaced contradictions and breaks in reasoning.
I asked Claude how long a refactor of an 800-line controller would take. Answer: half a day, a full day with tests. Same prompt, now with the mandate to actually do it: done in twelve to fifteen minutes, tests included. A factor of 30 between estimate and reality, on the same machine. The model estimates the way a 2023 senior engineer taught it to. Not the way it actually works.
MORE ON LINKEDIN →Per-seat means you pay for the tool to be there. Usage-based means you pay for it to run. Teams who had Copilot bundled as a dev benefit now have to measure AI output: which tasks go to the model, which output ships untouched, where am I paying credits for results I then correct. Exactly the conversations still missing from many AI pilots.
MORE ON LINKEDIN →Three npm supply-chain incidents in five weeks. What's new is the speed at which a compromised package slips into the workflow: AI coding agents pull and install dependencies faster than a human can read package.json. Allowlist your registries, run your own proxy, don't wave through install hooks, don't keep auth tokens in the same environment as an agent running npm install.
MORE ON LINKEDIN →
The roles are in the player card. What connects them: I build with my own money and my own name on the line, whether in an AG listed on the Düsseldorf Stock Exchange or in a crypto payment platform applying for its MiCAR licence. Beyond my own companies, I founded the AI Meetup Frankfurt, which at one point grew into the largest AI community in Europe.
Coding and self-employment both since school; my mother had to co-sign the first company registration at the town hall. At the Frankfurt e-commerce agency Sitewards I was CTO and partner with my own stake in the business, leading an engineering team of around twenty. Since 2015 I have founded and led my own companies.
My first production AI system ran in 2017 in the savedroid app. Since then every company of mine has had at least one production AI use case, several with regulatory relevance. The rest of the story is in the levels above.
With entwickler.press: "Security im E-Commerce" and "OWASP Top 10".
More appearances at Developer Conference (Hamburg), eShop Summit (Berlin), Magento Live (Munich and London), Meet Magento (Leipzig), PHP Unconference (Hamburg), Take Off Conference (Lille) and Webmontag (Frankfurt).
Former OWASP member.
Founded and grown into what was, at times, the largest AI community in Europe. Host role since handed on; the meetup continues under new leadership.
TO THE MEETUP →Mentor at Jugend hackt in Frankfurt: supporting young people who use code to improve the world.
TO JUGEND HACKT →Mail me at info@tobiaszander.de. Hacker, engineer and founder in Frankfurt Rhein-Main. If you're building something interesting, I want to hear about it.