studenti.rs
A student community in Serbia, online since 2008. I built it at my first job, bought it back in 2022, and rebuilt it this year.
- Registeredstudents
- 250k
- Documents
- 40k
- Online since
- 2008


A loop in five steps: a student uploads notes, it is checked and sorted, it becomes a page, search finds it, and another student opens it, which leads back to the first step. The library grows and stays clean without anyone reading uploads by hand. What still reaches me: uploads the final check is unsure about, held until I look; lifting a flag from an account; an alert when an upload gets stuck.
The whole picture
The library grows and stays cleanwithout anyone reading uploads by hand.
- StudentA student uploads notes
- Our AIIt is checked and sorted
- AutomaticIt becomes a page
- AutomaticSearch finds itGoogle + site search
- StudentAnother student opens it
What still reaches me: uploads the final check is unsure about, held until I look; lifting a flag from an account; an alert when an upload gets stuck.
A flowchart in three lanes: the student, the system and me. The student uploads a file. A model asks if it is study material; if not, it is refused on the spot. The student confirms the details. An automatic check asks if it is a copy; if so, it is binned, and a strike. A model asks if it is good enough: if not, it is binned; if it is not sure, I decide; if yes, it is published now or on a future date. Published, the student is rewarded and told.
From upload to published
Every upload is checked, sorted and published without waiting for a person.
- Student: Uploads a fileAny office format. Nothing is public yet.
- The system, a model, asks: Study material?Junk and spam stop here.If no, it is refused on the spot. If yes, it goes back to the student.
- Student: Confirms the detailsTitle, subject and summary are written for them.
- The system, automatically, asks: A copy?The same notes cannot be uploaded twice.If yes, it is binned, and a strike. If no, it goes on.
- The system, a model, asks: Good enough?Clear cases never need me.If yes, it goes on. If not sure, it comes to me. If no, it is binned.
- Me: I decideHeld until I look.
- The system, automatically: Publishes now, or on a future dateThe library grows a little every day.
- The end: Published. The student is rewarded and told.
A pipeline. A student uploads any office file: Word, PowerPoint, PDF or another format. It is converted to a PDF, its words are extracted as Markdown, and a web page is published. One branch sends the Markdown to the site’s search index. All 45,000+ older documents were re-run. 7,000+ were removed: duplicates, low quality, or material that should never have been public.
From upload to web page
One upload becomes a page anyone can find.
- Any fileWord, PowerPoint, OpenDocument, RTF or PDF: whatever the student has.
- PDFEvery file is turned into a PDF, so each one is handled the same way.
- MarkdownThe words are read out as plain text. A model writes the title, summary and keywords from it.
- PageThe free pages become a real web page that search engines and AI assistants can read and quote.
- Search indexThe site’s own search reads the Markdown copy, so students can search by meaning, not only by exact words. About 37,500 Markdown files.
- Re-runAll 45,000+ older documents were re-run. 7,000+ were removed: duplicates, low quality, or material that should never have been public.
Search, before and now. Before the rebuild, search matched titles and page text; the rest of a document was not searched. Now search reads whole documents and understands meaning, so a query can find a passage that says the same thing in other words. Before, 1 in 10 searches ended with the student finding what they wanted; 7 in 10 do today.
Search, before and now
Search now looks inside the whole document, not only at its title.
Before
Search matched titles and page text.
Now
Search reads whole documents and understands meaning.
1 in 10searches ended with the student finding what they wanted.
7 in 10do today.
A count. The old library was a mess: badly written titles, missing tags, a catch-all subject and documents under the wrong institution. Then 38,263 documents were sorted by institution in January 2026. 20,672 matched from the text, 101 by reading the whole PDF, 57 still ambiguous, 15,367 with no institution named in the text, 2,066 with none confirmed from the PDF.
Sorting the library
The old library was a mess: badly written titles, missing tags, a catch-all subject, and documents filed under the wrong institution.
38,263Documents sorted by institution, Jan 2026
| Group | Documents |
|---|---|
| Matched from the text | 20,672 |
| Matched by reading the whole PDF | 101 |
| Still ambiguous | 57 |
| No institution named in the text | 15,367 |
| No institution, confirmed from the PDF | 2,066 |
Before and after. Before the clean-up about 40% of documents were filed under the wrong institution; on 21 January 2026 none were, after 15,237 wrong links were removed. Filed under the right one: 35.7% before, 54% after. The other 46% never name an institution in their text and are left blank, not guessed.
Filed under the wrong institution
No document is filed under the wrong institution any more.
- Before the clean-up
- about 40%
- After, 21 Jan 2026
- 0
15,237 wrong links removed.
Filed under the right one
- Before the clean-up
- 35.7%
- After, 21 Jan 2026
- 54%
- The other 46% never name an institution in their text. They are left blank, not guessed.

A flowchart in three lanes: the student, the Grok bot and me. Every day a student writes to info@: questions, payment reviews. The bot reads and sorts the mail and asks if it is simple. If yes, it checks the facts and replies in the site’s own voice, and the answer is sent without me. If not, I answer what it hands me: refunds, goodwill, billing disputes and legal.
Customer service, done by Grok bot
Simple questions get an answer without waiting for me.
- Student: Writes to info@Questions, payment reviews
- Grok bot, a model: Reads it and sorts it
- Grok bot, a model, asks: Simple?Cancelling, deletion, “we’re checking”If yes, it answers on its own. If no, it hands the mail to me.
- Me: I answer what it hands meRefunds, goodwill, billing disputes and legal
- Grok bot, a model: Checks the facts before it writes
- Grok bot, a model: Replies in the site’s own voiceShort paragraphs, the informal “ti”
- The end: Answered and sent, without me.
