By Darsej Rizaj — Co-Founder & CTO, Star Labs
I’ve been running Star Labs for eight years. In that time I’ve probably lost count of how many senior developers have walked out the door, each one taking months of context with them. The architecture calls, the weird workarounds nobody documented, the “we tried that in 2019 and it broke everything” kind of knowledge. Gone overnight.
Replacing someone senior takes weeks at best. Real knowledge transfer? Never happens. You bring in someone new and they spend three months just figuring out where things are.
That pattern ate at me for a long time. Not just the turnover part. The whole setup felt wrong. How teams are structured, how knowledge flows (or doesn’t), how we’ve been using AI as glorified autocomplete. About a year ago I stopped complaining about it and started building something different.
What actually broke
Some context first. Hana Qerimi and I started Star Labs in Kosovo in 2017. We’ve got about 200 engineers now, a presidential medal from the President of Kosovo for our work in education, and a sister company called Digital School that teaches 40,000+ kids to code across 40 countries. We’re not some random agency. 70% of our devs come through our own talent pipeline, where we filter 500-600 applicants down to about five hires each quarter.
Even with all that. Even with 73% retention, which is honestly great for this industry. The same problems kept coming back.
Your best developer solves a gnarly bug at 11pm. That fix now lives exclusively in their brain. Maybe they mention it in standup. Probably not. Three weeks later someone else hits the same wall and wastes half a day rediscovering the solution.
Then there’s the silo thing. Front-end people refuse to look at infrastructure. DevOps won’t go near product code. QA only tests, never builds. I grew up treating engineering as one discipline. That mindset barely exists anymore.
And AI. God, the AI situation. Companies are out there buying ChatGPT Enterprise seats and calling it digital transformation. Meanwhile 80% of their devs are using it like a search engine with a chat interface. No memory between sessions. No accumulated context. The few developers who actually build up AI workflows and custom instructions? All of that sits on their personal machine. They leave, IT wipes the laptop, and the company’s “AI transformation” resets to zero.
So I built the Hive
I’m going to admit something embarrassing. The original idea came from Star Trek. Specifically the Borg. I’ve been a sci-fi nerd my entire life and the concept of a collective intelligence, where no single member holds knowledge that the group doesn’t also have, it just clicked for me. Obviously we’re not assimilating people. But the AI agents? Fair game.
The model I landed on is what I call a Constellation Team. Five people, five disciplines, everyone writes code. Not five frontend devs. Five different specialists who all ship.
The roles (I went full space theme, no regrets):
- The Architect — your senior dev
- The Navigator — product owner, but one who codes
- The Foundation — DevOps
- The Guardian — QA, but they build things too
- The Catalyst — AI and data
That last point matters. QA doesn’t just write test cases. They build features. The product owner pushes commits. Nobody gets to sit back and just manage or just review. Everybody ships.
But the team structure alone isn’t the interesting part. What connects them is.
Five layers inside the Hive
I’ll walk through these because they each solve a specific failure mode I kept seeing in traditional setups.
The Collective Brain. When our Architect burns an hour debugging a Docker networking issue, that entire troubleshooting path, the solution, the context around it, gets captured. Automatically. No one needs to write a Confluence page. When Foundation hits something similar two months later, the Hive already knows what worked. That’s hours saved on a single incident. Multiply that across a year.
Shared Memory. This one’s personal for me because it’s the problem that started everything. Normally, a developer builds up incredible context inside their AI assistant. Project structure, naming conventions, past decisions, all these micro-details that make them fast. Session ends. Context gone. Or worse, they leave the company and all of it vanishes.
In our setup, those AI memories sync to a central store organized by project. When the Guardian picks up work the Architect started, their AI isn’t starting cold. It has the full picture.
Real-Time Consciousness. Say the Architect is refactoring a payment service. Foundation, not knowing this, tries to redeploy the same server. The Hive catches it and pings both of them. No standup needed. No “hey did you see my Slack message from three hours ago.” It just works.
Session Persistence. Anyone who’s used Claude or ChatGPT for more than an hour knows the pain. Context window fills up, the AI starts forgetting things you told it twenty minutes ago. Our system logs every session. When one ends it gets compressed into a summary. If someone asks about something that fell through the cracks, the Hive searches those summaries. Worst case it can pull the raw session logs. Nothing gets lost. Ever.
Centralized Skills. Skills are reusable instructions. Deploy this way, review code like that, handle database migrations following this process. Normally each developer builds their own, if they bother. Most don’t. Those who do take it with them when they go.
We centralize all of it. A deployment skill, a code review standard, a frontend component pattern. It becomes company property, not personal preference. I genuinely think that in five years we’ll be valuing companies partly based on how deep their AI skill libraries are. Nobody’s tracking that metric yet. They will.
We killed Scrum
Not modified it. Killed it. No sprints. No planning poker. No retrospectives. No daily standups where half the team mentally checks out while waiting for their turn to say “no blockers.”
Here’s what replaced it:
- Demo meetings with stakeholders. That’s the only recurring thing.
- Crisis calls when something actually breaks.
- One chat channel. Not five channels with threads and sub-threads. One.
The part I didn’t expect: how happy people got. I keep hearing from team members that this is the best work experience they’ve had. I think it’s because they actually feel like they’re building things, not performing ceremonies about building things.
These teams don’t have fixed project scopes either. One week they’re building a client product, next week an internal tool, week after that they’re improving our infrastructure. The side effect is wild. Engineers who’ve been with us for years are discovering entire parts of the company they never knew existed. They’re getting context that normally only C-level people have.
I used to watch departments fight each other through layers of meetings and slide decks and escalation threads. That whole dynamic just… collapsed. When everyone sees the full picture, most of that friction disappears on its own.
What the numbers look like
We have six of these teams running right now. More coming.
| Metric | Traditional Team | Constellation Team |
|---|---|---|
| Team size | 10-15 engineers | 5 engineers + Hive |
| Monthly cost (EU equivalent) | €65,000-120,000 | €35,000-50,000 |
| Ramp-up time | Months | 2-4 weeks |
| Knowledge retention | Walks out the door | Permanent |
There’s one story I keep telling because it still blows my mind. A new developer joined a Constellation Team last quarter. No human onboarding. The Hive walked him through our projects, our architecture, our conventions, everything. End of day one he shipped a feature to production. His exact words to me afterwards: “I didn’t think that was possible on a first day.”
On the QA front. Automated testing at real scale was this white whale that every client wanted and nobody could deliver. What changed is the Hive generates massive test suites and then dynamically spins up infrastructure to run them. We’re talking 10-15 extra cores, additional RAM, all provisioned on the fly. Test runs that took hours now finish in minutes. First time we demoed it to the team, the room went completely silent. Someone eventually said “why has nobody done this before?” Good question honestly.
Where this goes next
I think about AGI a lot. Probably too much. But here’s my take: the LLMs we have today are impressive but they’re still isolated. They don’t know your company. They can’t remember what your team decided last Thursday.
When real AGI arrives, and it will, the companies that have a structured knowledge layer ready to plug into it will have an insane advantage. That’s what the Hive actually is underneath everything. A knowledge infrastructure that’s ready for whatever comes next.
We phase clients in over about twelve months. First three months is capturing institutional knowledge and assembling the team. Months four through six, the AI agents start generating documentation on their own, flagging patterns, suggesting solutions from past work. After that it compounds. Quality gates automate themselves. The system starts connecting dots between projects that humans wouldn’t have linked.
What I’m really building is a digital twin of a tech organization. A complete working replica of everything a company knows about its own technology. People can move between teams, they can leave, and the knowledge stays. It doesn’t degrade. It doesn’t forget. It actually gets better over time because every interaction feeds back into it.
Why I care about this
Strip away the tech and there’s something personal driving all of this. I’ve spent years watching founders and entrepreneurs, people who put everything on the line, get held hostage by their own teams. A key engineer figures out they’re irreplaceable and suddenly the power dynamic shifts. Salaries inflate. Knowledge becomes leverage. Companies start buying SaaS tools to fill gaps that their own people could’ve solved, if the knowledge hadn’t walked out with the last person who quit.
I don’t think that’s how it should work. The value should go both ways.
Constellation Teams don’t make engineers less important. The opposite actually. Five strong engineers producing the output of fifteen or twenty, working across the whole company instead of one narrow slice, having genuine fun doing it. That’s not a cost-cutting play. That’s what engineering should feel like.
We’re not writing white papers about the future of software teams. We’re shipping it. Right now. One constellation at a time.
Star Labs was founded in 2017 in Kosovo by Hana Qerimi and me. We’ve grown to over 200 engineers and received a presidential medal for our work in education and innovation. More at starlabs.dev.