Luka Abramović

Building a team that builds LLM wrappers. From sailor, to developer, to CEO of Posting about how I got here and where I'm heading next.

[Luka Abramović @lukaabram]

From 10 MB/Day of Internet onboard a container ship to running a 6-figure AI agency. The whole story is here

https://t.co/YKgnmmBZmM

[Luka Abramović @lukaabram]

Excited to share that my article for Fiverr’s Community Hub is live!

I wrote about using AI in real client work, what it takes to get quality results, and why AI is only as valuable as the thinking behind it.

Thanks to @fiverr for the opportunity!

https://t.co/Jw5it75H5O #FiverrCommunity

[Luka Abramović @lukaabram]

Guy who sells shovels upset that gold miners aren’t buying more shovels

[Bearly AI @bearlyai]

Jensen says he will be upset if he finds out his $500k engineer is *not* using at least $250k in tokens

[Luka Abramović @lukaabram]

If you ever feel bad about taking someone else’s product as inspiration for your own - don’t.

OpenAI’s CEO Sam Altman just thanked all the developers for writing code by hand in the days before LLMs, which enabled OpenAI to train ChatGPT.

Everyone takes inspiration from others who solved complex problems their own way brilliantly. We are all standing on the shoulders of giants.

Go back in time 4 years ago to the dark ages when there was no ChatGPT or other slop machines, and ask any developer how they solved problems. 99% will tell you they would search Stack Overflow and copy paste the solution. Hell, most tutorials I did when I started learning to code had me build a Reddit clone.

Just ask any UI/UX designer how they start their projects. They create a moodboard of other people’s designs they really like and go from there.

Nobody has an idea that wasn’t based on something someone else already made. The point is not to copy it blatantly, but to use it as inspiration, enhance it, and give it your own twist.

If you have an idea, the world is your oyster and it’s ripe for the taking. Just make sure you don’t end up with a $1.5B lawsuit like Anthropic

[Sam Altman @sama]

I have so much gratitude to people who wrote extremely complex software character-by-character. It already feels difficult to remember how much effort it really took.

Thank you for getting us to this point.

[Luka Abramović @lukaabram]

@FlorinPop17 @andrewzacker This is the highest value post I’ve seen from you. Great insights!

[Luka Abramović @lukaabram]

AI tools work great in English. Then you try Spanish and things break.

This is a pattern I keep seeing.

A team builds an AI content generation tool. They test it in English. Everything works beautifully. Then they start generating content in other languages and hit a wall.

English words randomly appearing in non-English output. Brand names getting mangled. Special characters breaking formatting. Prompts that worked perfectly in English producing inconsistent results in German, Spanish, or Croatian.

And the thing is, this is not one bug. It's a whole cluster of issues that show up together.

We're seeing this firsthand right now. Most of the issues our teams are dealing with are tied to non-English content generation. And this is not unique to any single LLM. It's a gap we noticed across the board.

Most large language models are English-first. Training data, benchmarks, fine-tuning. So when a company reads "supports multiple languages" on the spec sheet and assumes that means "works equally well in multiple languages," they're in for a surprise.

The worst part? Most teams discover these issues after launch. Because nobody thinks to stress-test the AI in Portuguese or Dutch during the build phase. At least not as diligently.

So if you're scoping an AI content tool for a multi-language market, here's my honest advice. Budget 20-30% more what you planned for non-English QA. Get native speakers involved during prompt engineering, not just final review. And build evaluation pipelines for every target language from day one.

The companies that have figured this out will have a real competitive advantage for multi-lingual content generation.

[Luka Abramović @lukaabram]

Stop building features. Fix the ones you have.

If you're building an AI product and every feature kind of works but nothing works great, adding more features is the worst thing you can do.

The gap between "demo-ready" and "production-ready" is enormous, especially with AI. A feature with too many false positives isn't a feature, it's a liability. A tool that works "kind of okay" is something your team stops using within two weeks.

I know it's tempting. Every stakeholder wants the next thing. The roadmap is screaming for more. But before you add feature number four, ask yourself if features one through three are actually reliable.

Shipping three mediocre AI tools will always be worse than shipping one great one. Sometimes the most productive sprint is the one where you build nothing new.

[Luka Abramović @lukaabram]

No-code worked great. Until it didn't. At least for us.

Now this post might ruffle some feathers, but I've had multiple people come in wanting complex automations and insisting on Zapier or Make. They've seen the demos, the marketing looks great, and I completely understand the appeal.

So we tried it. Multiple times, with multiple clients. And honestly, for us it was a complete mess.

The pattern was always the same. Everything looks promising at first, but then edge cases multiply, error handling becomes a nightmare, debugging is completely opaque, and the team realizes they need custom code anyway. But now weeks are already gone. Also the development time seems like it's slower. Yes, the deployment is easier, but the deployment is not the hardest part with custom projects.

Now, I want to be fair here. There are agencies out there that do no-code incredibly well, even for large and complex projects. They've built real expertise around these tools and they deliver great results with them.

But after trying it ourselves multiple times, we made the decision to go back to custom development for all use cases. And the reason is simple. We're better at custom development, we're faster writing code, we know the tools, and we know the process like the back of our hand.

That's not a knock on no-code. It's just me being honest about where we do our best work.

So if you're really insisting on no-code automation platforms, more power to you. But I feel today with the help of AI coding agents, everyone can be fast and productive building custom code solutions.

[Luka Abramović @lukaabram]

Have the same confidence that this FC Augsburg app has when it tells users the time (incorrectly).

It seems like a test function slipped into production or somebody mixed up environment variables while testing.

Anyway, when building an app, always double check what's going live and in what environment you're currently testing. As funny as it is, these types of things confuse users and lose their trust.

[Harry @Owens_EFC]

downloaded the FC Augsburg app for digital tickets cos im going next week and it randomly sent me a notification at 19:08 simply saying "it is 19:07"

[Luka Abramović @lukaabram]

Your developer isn't slow. You just hired a thinker when you needed a shipper.

Pre-product-market-fit? You need someone who moves fast and ships scrappy.

Scaling something that works? You need someone who architects for the long term.

[Luka Abramović @lukaabram]

A fintech founder told his investors he'd never hire a single support agent.

Not as a long-term goal. As the foundation of the entire company.

Most companies treat AI like a bolt-on. They take existing processes and try to squeeze out 20% efficiency gains. That's optimization.

This founder is building a company that structurally cannot exist without AI. The cost structure, the margins, the entire business model assumes AI handles operations at scale with minimal human involvement.

But this is fintech. People's money.

What happens when the AI hallucinates financial advice and there's no human backstop? What happens when a customer has a complex dispute and there's nobody to escalate to?

There's a difference between building lean and building fragile. The companies that figure out where that line is will be the ones that actually survive.

What do you think, visionary or reckless?

[Luka Abramović @lukaabram]

Giving AI more context can make it hallucinate more, not less.

We learned this the hard way. We were building a multi-step AI pipeline and decided to tell the AI what type of output we were producing during the research phase. More context, better results, right?

Wrong. The AI started skewing results toward confirming what it expected to find. Basically hallucinating data that fit the desired output. Confirmation bias, but inside an AI pipeline.

It's like telling someone which wine is expensive before a blind taste test. You've corrupted the data.

So here's the framework we use now:

Research steps should be context-naive about the final output. Just gather information. Analysis steps can know the output format. Generation steps get full context.

The most important question for each step in your AI pipeline isn't "what should this step know?" It's "what should this step NOT know?"

That one shift made our outputs significantly more reliable. If you're building multi-step AI workflows, map out each stage and carefully scope what context goes where. The order in which you introduce information matters just as much as the information itself.

[Luka Abramović @lukaabram]

Gold

[Trung Phan @TrungTPhan]

Traditional SaaS is dead.

I asked Claude to vibe code a Docusign replacement.

6 hours and 450k lines of code later it built a drag-and-drop PDF signer, initials stamps, and a very satisfying “signature complete” animation.

So, the first 45 contracts we used it for auto-selected governing law at random. Every agreement is void. We lost $473K in bookings and I have in-person courts dates in 18 countries over the next week.

But, man, was the dynamic completion checkmark graphic sick.

[Luka Abramović @lukaabram]

"My CS friends said it wouldn't take long."

The most expensive sentence in software.

I hear this on almost every discovery call. And I completely understand why it happens.

A founder asks a CS friend how hard something would be to build. The friend says "not too complicated." And now that estimate becomes the baseline for every future conversation.

The friend isn't wrong though. They're just answering a different question.

"Could a competent developer build something like this?" is a very different question from "What does it take to build this to production quality with edge cases, integrations, security, and real users?"

Let's say your friend estimates 2 to 3 weeks. That might be right for a prototype. But add payment processing, authentication, admin dashboards, testing, and deployment, and you're looking at a completely different scope.

When this comes up on a call, I like to acknowledge it, then walk through all the pieces that make something production-ready. Usually within a few minutes, the founder sees the full picture on their own.

The goal is to upgrade the conversation from "how long does the coding take" to "what does it actually take to launch this thing."

[Luka Abramović @lukaabram]

Azure will give your AI startup free credits. Three of our clients got approved in days last year.

1. Microsoft for Startups: $5k
2. AWS Activate: $1k
3. Google for Startups Cloud: up to $200k

Before you spend a dollar on cloud, check these programs.

[Luka Abramović @lukaabram]

The absurd thing about this bit is that Brits are doing just that to every tourist spot in Europe with shitty pubs whose sole purpose is catering to British tourists.

[Aɴᴛ @AntSpeaks]

British comedian Pat Smith brilliantly points out the hypocrisy surrounding the subject of mass migration in the UK.

[Luka Abramović @lukaabram]

@GlupeusMaximus @lxstlosty Let’s talk. Do you know anything about the different cultures, religions, and people in Yugoslavia? If you did you would realize your comment doesn’t make any sense

[Luka Abramović @lukaabram]

@JRbwaker @lxstlosty “Centuries” ago is 100 years ago actually. Not really that long to completely forget. The impact of it are still felt today globally with the mess that was left everywhere

“If you see two fish fighting in water, you can be sure an Englishman passed by five minutes ago”

[Luka Abramović @lukaabram]

@AntSpeaks Sure. However, the Brits did care about extracting wealth and resources from half the world for their own gain.

Don’t pretend like you’re the victims in this scenario when your country is the biggest reason for celebrating independence day worldwide

[Luka Abramović @lukaabram]

Where can I find the leaderboard?

[Hiastra @Hiastrax]

#8 Rank CookieKing just THRONEMAXXED #2 ASU Frat Leader right after Clavicular reclaimed #1 on the FrameMog Z leaderboard 😭📊

This could trigger an instant RANKGOON and force a full MOGOFF for the crown 🏆