Most stories about AI and engineering start with someone who already ships code for a living. This episode of Very Good Engineering comes at it from a different angle. Host Jorge Coca sits down with Peter Trost, a senior Flutter engineer who studied machine learning, fell for app development, and is now using AI to build a property management startup largely on his own — production-quality software with a fraction of the code he would have hand-written a few years ago.
About Peter Trost
Peter is a senior Flutter engineer at 1KOMMA5°, one of Germany’s fastest-growing clean-energy companies, where he works on the software behind home energy systems. On the side he is building a property management startup, leaning on AI to do the work of a much larger team. His background is a little unusual for a mobile engineer: he came up through computer science with a focus on artificial intelligence and machine learning, which gives him a clear-eyed view of what today’s AI tools are actually good at, and where they still need a human.
From Machine Learning to App Development
Peter’s path started in AI research, not app development. During his studies he expected to spend his career training models. What he found instead was that the day-to-day of machine learning — cleaning data, wrangling datasets, waiting on training runs — did not give him the fast, tangible feedback he wanted. Building applications did. He could make something, see it on screen, and hand it to a person to use.
That pull toward user-facing work led him to Flutter, and the developer experience is what made it stick. Hot reload let him change code and see the result almost instantly. The widget inspector made it easy to understand and debug a running layout. Coming from Python and Java, he found Dart intuitive and genuinely enjoyable to write. For someone who valued a tight feedback loop, Flutter closed the gap between an idea and a working screen.
Building a Property Management Startup with AI
Peter’s current project takes aim at property management, an industry still weighed down by manual inspections, paperwork, and slow back-and-forth between managers, owners, and tenants. His goal is to strip out the busywork so the people doing the work can focus on the property, not the admin.
Two features show the shape of the product:
- Voice-first inspection reports. A property manager walking a unit can simply talk, describing what they see as they go. The app transcribes the audio and structures it into an actionable report, turning loose observations into work orders without anyone stopping to type.
- Digital checklists. Guided checklists walk an inspector through a property step by step, so nothing gets missed and every inspection produces consistent, comparable records.
The through-line is the same one that drew him to Flutter: take a slow, error-prone manual process and give the person a faster, cleaner way to do it.
AI as a Development Partner
The most striking part of Peter’s workflow is how he treats AI — less as an autocomplete and more as a collaborator. He started, like many people, generating snippets with ChatGPT and Copilot, and has since moved toward agentic tools like Claude that can take on larger, multi-step chunks of work.
Two of his framings stand out. “AI helps me identify blind spots,” he says: the model catches cases and considerations he might have missed on his own. And “AI amplifies good practices.” If you bring clear structure, tests, and conventions to the table, the tools make you faster. If you bring mess, they amplify that too. The quality of what comes out tracks the quality of what goes in.
He also closes the loop between writing code and checking it. Using LeanCode’s Marionette, an MCP server that lets an AI agent drive a running Flutter app — inspecting the widget tree, tapping, typing, scrolling, and taking screenshots — the agent can actually exercise what it just built and see whether it works. That turns his core principle, “iterate until the goal is reached,” into something the tools can do alongside him rather than something he has to babysit by hand.
Reaching Production Quality with Minimal Code
“Minimal code” is easy to misread as “low quality.” Peter’s point is the opposite. The code he ships still has to clear a production bar — it runs in a real app for real users — he is just writing far less of it by hand. The leverage comes from pairing a framework with excellent ergonomics, Flutter, with AI tools that handle the repetitive scaffolding, then keeping tight control of structure, testing, and verification so the output holds up.
That is where his machine learning background pays off. He knows these models are probabilistic, not magic, so he leans on solid engineering practices — clear specs, tests, runtime checks — to keep them honest. The AI does more of the typing; he stays responsible for the judgment.
Bringing AI Workflows into the Enterprise
The same instincts carry into his day job. At 1KOMMA5°, Peter has helped introduce AI-assisted workflows inside a larger engineering organization, where the questions differ from those a solo founder faces. Adoption is as much about people and process as it is about tools: giving teams patterns they can trust, being honest about where the tools help and where they don’t, and building the guardrails that make it safe to move faster. His advice skews practical — start where the pain is real, keep a human accountable for quality, and let results, not hype, decide how far you push.
Balancing Focus and Exploration
Running at a startup while holding a senior engineering role forces a constant trade-off between focus and exploration. Peter is candid about the entrepreneurial tension that comes with it: every sign of traction is also a pull on your attention, and the discipline is deciding what to chase and what to leave alone. New AI capabilities land almost weekly, and the temptation to try every one of them is real. His answer is to stay anchored to the goal and let that decide which experiments are worth the detour.
Key Takeaways
- Optimize for the feedback loop. Peter chose Flutter for hot reload, the widget inspector, and a language he enjoyed. Fast feedback is what keeps momentum.
- Treat AI as a collaborator, not a vending machine. It surfaces blind spots and amplifies whatever practices you already have, so bring good ones.
- Close the loop with runtime checks. Tools like Marionette let an agent actually run the app it wrote, turning “iterate until it works” into part of the workflow.
- Minimal code still means production quality. Less hand-written code, same bar. Structure, tests, and verification are what make that true.
Conclusion
Peter Trost’s story is a good reminder that the lines between AI researcher, engineer, and founder are getting blurry. With the right framework and the right tools, one person with a clear goal can build and ship real software that used to take a team. His advice for anyone eyeing that path is refreshingly plain: pick tools with a great feedback loop, bring your own good practices, and iterate until you hit the goal.
