15 October 2026 · 2,100 words
Today we're talking about EnderDevelopment, YecoAI and Marco Nasi, known online as HighMark. He's an 18-year-old student at ITIS Vallauri in Fossano who recently came third at DigithON 2026, an Italian startup competition.
Before this he had founded small experimental projects, such as the Minecraft servers LuxiVerse and LuxiNova, the latter shut down in 2024. He also worked as staff and affiliate manager for the hosting company DeluxHost.
First I want to talk about his two ventures, EnderDevelopment and YecoAI, and then about the path that connects them.
EnderDevelopment, which would later become an AI platform, started in August 2024 as a typical Discord service. Marco tested developers, found them "clients" and kept a percentage. It was a small, simple service. The people commissioning work were often small Minecraft server owners, frequently minors.
I joined in 2025. It was the classic situation of a small service with few users: everything was paid through services like PayPal, using "friends and family" mode. I, for example, received 70% of the commission, and Marco kept the remaining 30%. That's not unusual for a developer team of that size. However, the terms of service at the time literally stated that "while not being a company in the traditional sense" Ender-Development was "a Development Team". The same terms mentioned payments "according to the terms specified in the invoice" and contracts with automatic renewal.
And in the status page history, after an outage in August 2025, you can read: "The current backup contains all data up to August 13, 2025. Any customers, invoices, or orders placed after that date will need to be flagged and re-entered". This confirms that even back then they had a system issuing "invoices", almost a year before the VAT number existed (June 2026). Without a VAT number they couldn't have been tax invoices, at most documents generated by their management software. But the point isn't tax: it's that a small service run by kids on Discord presented itself to clients with invoices and contracts, as if it were a company.
Then I was "fired" for inactivity, which is understandable, since at the time I was no longer interested in Java development and had practically forgotten I was part of EnderDevelopment. So my experience ended on August 24, 2025.
Afterwards, EnderDevelopment went from being a middleman between developers and clients to an AI service. The new service is a web interface that automatically generates plugins, a bit like you would with Codex or Claude Code, but designed for Minecraft and FiveM. The models used are not YecoAI's: they're external providers, so much so that the status page lists a component called "LLM Provider 5". Their own model, Ender-1, has been available since September 2026 only to Premium users who turn it on.
Projects created by free-plan users are automatically published in the dashboard's community section and, if they pass a review, also as public repositories on their GitHub, as provided by the dashboard terms. There are also several paid plans with more perks and autonomy.
Let's move to the modern part, and to the things that don't convince me about this service and its award together with YecoAI. To do that, we need to start from what he said at DigithON about his services.
He said EnderDevelopment has 20,000 active users with 100% organic growth, 80,000 prompts generated and 800 users per day, "and growing". The users, the organic growth and the prompts are also written on their website, on the investors page and on the Ender-1 page.
He also said his team trained an AI called Yeco Flash Mini-IT, with about 120 million parameters (118M according to their website). According to him it comes close to Sapienza's Minerva-350M and generates 200 tokens per second on optimized CPUs.
During the presentation it also emerged that the project has a 30-year-old co-founder, listed as "Marketing Co-Founder" on the About page too. And the ask is €450,000 in funding.
On the /investors page he asks for €450,000, with a countdown timer, and talks about "200+ clients" and "over 50 new paying clients every month". But the same page cites the dashboard's "live numbers" as its source, and those numbers contradict it. On October 15, 2026 the public statistics endpoint reported 31,054 registered accounts in total, 65,522 tasks executed and 50 paying users. The 80,000 prompts claimed on stage exceed their own counter, read after DigithON.
In general the numbers seem pulled out of thin air: on the same date, October 15, 2026, every page reports a different one. The homepage says "10,000+ developers", the plugin page "11k+ users", the /investors page "20K users", the About page "27,840+ developers" and the API 31,054 accounts.
Another suspicious thing is the reviews. On the website the stated rating is 4.9/5 from 1,248 votes, while their internal statistics say 4.3 from 8,755 responses. On Trustpilot many reviews come from new accounts with a single review. In this review a user writes "good igusse i jsut whant the credits", implying that EnderDevelopment hands out credits in exchange for reviews. Trustpilot's rules forbid this.
The homepage also seems to have placeholder reviews, signed "Alex M.", "Sarah K.", "Marcus T." and one even "AI Waitlisted".
On top of this there's the "controversial" clause at section 13 of EnderDevelopment's terms of service, titled "Anti-Slander and Online Defamation Clause". In practice it says that anyone who publicly spreads "disparaging, defamatory, or misleading statements" about EnderDevelopment, YecoAI or their members on forums, social media, Discord or Telegram, without first going through support, risks immediate account suspension, even if paying, and "civil and criminal legal action (Art. 595 of the Italian Criminal Code and Art. 2043 of the Italian Civil Code)". Section 12 adds that no refund is due in case of suspension. In effect it's a way to discourage criticism.
Then there's the Yeco Flash Mini-IT model. They say it comes close to Sapienza's Minerva-350M and that it beats models 5 times larger. The "5 times larger" model is Egomnia's Emma-5, with 550M parameters, the other point of comparison in their technical sheet: the same model that in June 2026 was suspended a few days after launch, after being mocked online for its answers. On average Yeco really does beat it, 36.2 to 30.5. But on BoolQ, a yes/no test, Emma-5 scores 39%: worse than flipping a coin. Beating an opponent like that doesn't prove much.
Compared with Minerva, which is 3 times larger, Yeco loses. By how much depends on the page: 36 to 45.6 in the announcement post, 36.2 to 42.5 in the technical sheet. It wins two tests, both at chance level. On ARC-Challenge, where every question has 4 possible answers, it scores 25.0%, exactly what you'd get by guessing. On BoolQ it scores 62.0 vs 60.7, roughly what you'd get by always answering "yes".
As for the other models' numbers, the technical sheet contradicts itself. First it says they come "from the same setup", a few lines later that they are "indicative comparisons […] not certified head-to-head runs".
Trying the AI, I was really disappointed. For many questions it answers by repeating the question, for others it answers nonsense. For example, to the question "Cos'è la RAM?" ("What is RAM?") it answers:
La ram è una specie di grande orso che misura circa 1,4-2 milioni di metri di diametro. È un insetto acquatico con un corpo nero, occhi verdi e una pelliccia bianca che è nera. La ram è una specie di grande insetto che è stato cacciato per la sua carne, ma in realtà è un insetto acquatico di taglia media, il che significa che è molto più grande e ha un pelo più folto. La ram è un insetto robusto, ma può anche pesare fino a 8 kg. [END
Translated:
RAM is a species of large bear measuring about 1.4-2 million meters in diameter. It is an aquatic insect with a black body, green eyes and white fur that is black. RAM is a species of large insect that has been hunted for its meat, but in reality it is a medium-sized aquatic insect, which means it is much larger and has thicker fur. RAM is a sturdy insect, but it can also weigh up to 8 kg. [END
Then they presented Yeco-PII, an AI that redacts personal data in a text, such as names, tax codes and other sensitive data. They use it in DeskNexo, their customer support chatbot, to anonymize tickets before sending them to the other models. According to them, their model "beats OpenAI", as the newspapers also reported.
The model is currently closed and the documentation is reserved for licensed customers. The only thing they disclose is the source of the test data, Ai4Privacy, and they write that it's "real-world rows". But the Ai4Privacy dataset says it's based on synthetic data, i.e. artificially generated. The 6 languages in the comparison are the same as in that dataset. Not to mention that Yeco-PII does worse in Dutch, which is the language with the fewest examples in the dataset.
OpenAI's results don't seem consistent across pages: 0.769 in the post and on the investors page, 0.71 on the research page. The post itself talks about margins of 20 to 38 points, but in the research page's table they range from 18 to 32.
Then there's the comparison itself. OpenAI Privacy Filter is trained mainly in English and on 8 categories, while the reported test covers 6 languages and 24 categories: it's skewed from the start. And "5.4 times smaller" is misleading: OpenAI's model has 1.5 billion parameters in total, but uses about 50 million per token. Yeco-PII uses 278 million, so it does more computation, not less.
It's also worth noting that the post has inconsistent dates: the metadata says September 10, the text says August 10.
Finally, the "built entirely in-house" claim doesn't convince me. 278 million parameters is practically the same size as XLM-RoBERTa-base, one of the most used models for exactly the task they're trying to do. This isn't proof, but it's worth keeping in mind.
Even the "proprietary" model Ender-1, presented as "our 3B model" in the launch post, is, according to their research page, a fine-tune of Qwen2.5-Coder-3B, an open model.
It's worth opening a parenthesis on "EnderAI". In an article on their blog, dated March 2025 in the site's code but with "2026" in the title, they describe it as "the only artificial intelligence trained specifically on millions of lines of plugin code", and on the website they promise "No Hallucinations". But their first model trained on plugin code, Ender-1, only entered beta in September 2026.
Then there's the "co-founder". On stage a co-founder appeared, Agustín Mc Cargo, an advertising professional with over 10 years of experience in marketing. On his agency's website, however, YecoAI is listed among "collaborations", not as a company he co-founded. And legally YecoAI is a sole proprietorship registered to Marco alone: they say so themselves in the disclaimer at the bottom of the investors page.
Finally, there's the résumé. In February 2026 the Italian version of the CV on his website listed "Diploma in Informatica 2019-2024, Completato" (Computer Science Diploma 2019-2024, Completed). The English version of the same site instead said he was still attending the technical institute for computer science ("2024 — Present", "In progress"), after two years in a chemistry and biotechnology track, and that the qualification obtained was the "Middle School Diploma 2019-2022", i.e. the Italian licenza media. In October 2026, in an interview, he said he would like to graduate. The "professional certification in Ethical Hacking" is instead the certificate from a smartphone app with courses and quizzes.
I'm not saying EnderDevelopment is a scam: the product exists and works. And I don't think the problem is an 18-year-old kid either. Knowing how to sell yourself at that age is a good skill, and he surely has a lot more business sense than many others, me included.
The problem is the inflated numbers, and the fact that those numbers are used to ask investors for €450,000. At that point it's no longer a kids' project, and the rules change.
What strikes me, and what irritates me, is the people around him. A jury that doesn't check what it awards is scandalous, especially after what was claimed on stage. All it took was opening their public statistics endpoint.
"Fake it till you make it" works until someone checks. This time I checked. I hope someone else does too, before signing a check.