Tomáš Hromada

The biggest advantage I had in my twenties wasn't talent or connections. It was my naivety. I would work with early-stage startups where everyone seemed to get promoted fast. Someone would be a junior recruiter for two years, then become a senior. A year or two later, they'd become a lead. Then head of department. I genuinely thought that was how careers worked everywhere. That you could climb from junior to head of department in just 5-6 years. So when someone offered me a senior title after two years, I said yes. When a lead role opened up after three years, I took it. When a head of department position came up, I went for it. I kept saying yes to titles I wasn't actually qualified for yet. Now, after 10 years in recruitment, I understand what a true senior-level person or true head of department actually is. It's someone with deep expertise. Who knows how to handle complexity and can mentor others because they've already made the mistakes. That's not what I was when I took those titles. But because I didn't know I "wasn't ready," I just did the work anyway. I was always a step behind the title I held. Learning on the job, figuring things out as I went, working hard to catch up to what the role actually required. Startups gave me the space to take on responsibilities I had no business handling yet. I learned fast. I made mistakes and fixed them. I grew in real time. The gap between what I thought I could do and what I actually could do forced me to close that gap quickly - or fail. Ten years later, I'm running Satellite Talent with a much clearer understanding of what real expertise looks like. But I only got here because I didn't know enough to be scared of the opportunities in front of me. Here's what I tell people early in their careers: If you feel slightly unqualified for an opportunity - that might actually be a good sign. That discomfort means you're growing.

118

22 Comments

The biggest problem with AI-driven software development is that the English language is inherently ambiguous and open-ended. Even when given relatively clear instructions, AI often misunderstands requirements or invents details that were never explicitly requested. This makes it difficult to achieve the desired outcome. As a result, you frequently need to provide additional prompts, corrections, and refinements to guide the AI toward the intended solution. Furthermore, AI often tends to add new code rather than modifying existing code. Over time, this can lead to an unnecessarily large and complex codebase, making it harder for both programmers and AI systems to understand, maintain, and fix issues in the future. What if there were a language that was more precise, produced the same output every time it was executed, and generated predictable results? If such a language existed, it could significantly improve software development. Perhaps it already does: programming languages. :)

6

8 Comments

We made a bet 5 years ago that turned out to be accidentally genius: When we started Checkly, we believed monitoring should be code that you can review, version, and test, just like the rest of your system. Developers liked this because it matched how modern teams already work. That was the main idea. Monitoring as Code is like Infrastructure as Code; automation fully embedded in your SDLC. Then AI agents arrived. And suddenly, what started as an engineering choice became a big strategic advantage. Here’s why: AI agents are very good at working with code. They already understand tools like Playwright, TypeScript, and OpenTelemetry. These are open standards, so agents already have context for them. But asking an agent to navigate a dashboard, click through a UI, or understand a visual workflow builder is much harder. Teams that built monitoring around GUIs (ClickOps) and drag-and-drop interfaces are now trying to figure out how AI agents will use their products. We don’t have that problem, because code was always the interface. Today, we even have an engineer working full time on improving our CLI for AI agents. It now gets more than 130,000+ downloads weekly and growth keeps increasing. We weren’t trying to predict anything. We focused on strong and open fundamentals, and they paid off.

51

5 Comments

I decided to build one product every day in March. I shipped 16 open source projects in the first three weeks. Then I spent the last week building something different: an autonomous software factory. My website has had more than 300 commits in the last week. I made maybe 5-6 of them. The factory is an autonomous system that manages aikaara's entire website. It's not a CI/CD pipeline or a cron job that runs tests. It's an AI system that decides what to build, builds it, tests it, and ships it. Here's how it works: The Planner runs every 5 minutes. It audits the live site, reads the vision doc and current backlog, then generates prioritized coding tasks. It deduplicates against existing work and commits the updated backlog. The Factory picks up tasks, spins up a coding agent, and gives it full context: codebase docs, project specs, constraints. The agent writes code, runs the build. If the build fails, it rolls back. If it passes, it commits, pushes to production, updates the backlog, and documents what it did. No human in the loop. 182 factory commits. 145 tasks completed. It started with a full About page (JSON-LD schema, SEO, the works). Then it moved to GA4 conversion event tracking. From there it picked up blog CTAs, email capture, and table of contents generation. It hardened security across the site (noopener noreferrer on all external links), enhanced the contact form with privacy assurance, added breadcrumbs, meta tags, and og:image support. It turned the 404 page into a conversion opportunity. Then it went after performance: lazy loading, code splitting, preconnect hints. Each commit includes build verification. Each task gets documented in a codebase log. The system maintains its own institutional memory. The interesting part isn't any individual feature. It's that the system identifies what's missing, prioritizes it against business goals, and executes without being told. Today it tried to build the same navbar dropdown 3 times because the planner kept generating title variations that bypassed dedup. So I fixed the dedup logic. That's the real work now. Improving the system, not building the website. I spend my time on the factory. The factory spends its time on the product. This is what changes when AI agents actually work end to end. The agent decides what code to write, writes it, validates it, ships it, and documents it. While you sleep. #softwarefactory #aikaara #ai

26

2 Comments

Just attended "Agents on the Autobahn: How Grammarly Builds LLM-Powered Features at Full Speed" at Grammarly #Berlin. Expected AI hype. Got a masterclass in software craftsmanship instead. Vladyslav Bozhko and Yevhenii Sabanin explained their success and it wasn't Agents, AI or automations. They did classic software engineering excellence - they did it well, they did it fast but it worked because it is engineered to work! Architecture that scales: • Mono repo with clear package boundaries • Small, independent components • Incremental releases • Feature flags Process that delivers: • Cross-functional teams, small pods and with functional alignment beyond the teams • Automated pipelines with built-in testing • A/B testing as standard practice Decision-making that matters: • Data-driven at every level • Active UX research and customer feedback loops • Direct user interaction informing development The LLMs? Just another tool in the stack. One that adds user value and enables the teams but not a magic bullet! The real story: You can't skip fundamentals. Not when integrating databases. Not when calling APIs. And definitely not when working with probabilistic AI systems that need tight guardrails. Solid engineering practices. Clear team structures. Fast feedback cycles. Customer-centric development. These principles don't change with the technology stack. They become more critical. This is how you ship reliable features at scale - whether you're building with LLMs or anything else. The discipline is the differentiator. #SoftwareEngineering #AI #LLM #EngineeringExcellence #Grammarly

31

11 Comments