Parallax
Contributors
Parallax is one of the two projects our team built at Junction 2026, the hackathon in Espoo from 13 to 15 November 2026. We entered it for MariaDB's challenge Wikipedia Through Time. Out of about 800 projects it made the top 5 of the main event, so we pitched it live to the grand jury in the final.
Most people read Wikipedia in one language and never notice that the same article says something else in another one. Parallax puts the Finnish, Swedish, German, English, Russian, Estonian and Ukrainian editions side by side, statement by statement. The Winter War is a good example. The Finnish article says the Soviet Union staged the shelling of Mainila. Since August 2009 the English one has only said that an incident was reported and in July 2014 the Russian one dropped the part about the Soviet side staging it.
What it does
For an article, Parallax lists every statement that another edition tells differently and marks the number, date or word where they part ways. Click one and you see it in all seven editions, with the date each of them added, reworded or removed it. A date picker cuts every edition at the same day in the past. On 1 June 2008 the Winter War had 4 differences, on 14 November 2026 it had 24. Another page collects what one edition removed while the others still have it.
It also checks what a chatbot tells you. Parallax splits the answer into statements and looks for each one in every edition over its whole history. A statement comes out as backed, disputed (with the year an edition removed it) or in no edition at all. When we asked Gemini how the Winter War started, it gave us 17 statements. Eleven were backed and five were in no edition. One was disputed because the Estonian edition removed it in 2009.
How it works
All 875,572 revisions sit in MariaDB 12.3 system-versioned tables, next to a bge-m3 vector for each sentence. A single query combines FOR SYSTEM_TIME AS OF with vector search and finds the closest sentences in the other editions on a given day in about 5 milliseconds per article. Gemini 3.8 Flash then decides for each pair whether it is the same statement, marks what differs and translates it into English. When Gemini is too slow the vector similarity decides and the statement says so.
Any article with a Finnish version can be loaded from the start page and is ready after one to two minutes. During the hackathon it ran on a Verda server with Docker Compose and Caddy.
We were a team of four and the code is open source under the MIT licence.