Valira AI

New blog post: 𝗵𝗼𝗺𝗼𝗴𝗿𝗮𝗽𝗵𝘆 𝗶𝗻 𝗮 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲 📐 In one of our recent projects, we ran into an interesting challenge: estimating the speed of a moving object using only video. Our solution? Homography. Homography is a geometric transformation that relates points between different views of a scene. In practice, it lets you map image coordinates onto a consistent real-world plane, making it possible to measure distances, trajectories, and ultimately speed from a single camera setup. We’ve put together a short write-up explaining the concept and how it shows up in real computer vision systems. Read it here: https://lnkd.in/d4XmtwpU Author: Blaž Pridgar #ComputerVision #AI #MachineLearning #Engineering

On Wednesday, we held a workshop on AI agents and the Claude ecosystem. Together, we explored practical use cases, demonstrated how AI agents can be integrated into everyday workflows, and discussed ways organizations can optimize processes and increase productivity with these emerging technologies. A big thank you to Slovenian AI Factory - SLAIF for the invitation and for bringing together a great group of participants. We enjoyed the discussions and look forward to continuing the conversation around the future of AI.

Our third and final blog post on LES (lestvica embeddingov za slovenščino) is live! This time, we round off our Slovene-focused embedding benchmark series with practical guidance on 𝘀𝗲𝗹𝗲𝗰𝘁𝗶𝗻𝗴 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 for your specific task. 🔍 👉 Check it out here: https://lnkd.in/djzqHrwD Author: Neli Čatar #NLP #Embeddings #Slovenian #MachineLearning #Benchmark #LES #DataScience

Yesterday we attended the Data Science Project Competition Finals where the top 5 teams of this year's competition presented their work from the past 5 months, collaborating with industry mentors on real-life projects. 🏆 We are proud to announce this year's winning team was mentored by our colleague Andrej Miščič! Tjaš Ajdovec, Gal Dvoršak, and Denis Nedić developed 𝘀𝗹𝗢𝗖𝗥 — the first comprehensive Slovenian-language OCR benchmark. While optical character recognition has advanced rapidly in recent years, evaluation remains overwhelmingly English-centric, leaving morphologically rich languages like Slovenian without a standardized way to measure performance. slOCR addresses this gap with a diverse, human-annotated corpus of 632 real Slovenian sources spanning six categories: from clean academic documents to degraded historical scans and handwriting. The team evaluated 20 systems, ranging from traditional pipelines and OCR-specific transformers to open-source vision-language models and proprietary APIs, using a metric suite tailored to Slovenian's structural and orthographic specifics. Key findings: 📌 The gap between model classes nearly vanishes on clean text but widens sharply on degraded and handwritten inputs, where proprietary models prove most robust. 📌 Gemini 3.1 Pro performed best overall. 👉 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗿𝗲𝘀𝘂𝗹𝘁𝘀 𝗮𝗿𝗲 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝗱 𝗽𝘂𝗯𝗹𝗶𝗰𝗹𝘆 𝗼𝗻 𝗛𝘂𝗴𝗴𝗶𝗻𝗴 𝗙𝗮𝗰𝗲: https://lnkd.in/dCYkAv39 Congratulations to the team on a well-deserved win! 👏 And a big thank you to: - DataScience@UL-FRI for organizing such an excellent competition, - Domen Vreš and Erik Štrumbelj for co-advising, - Špela Arhar Holdt and Iztok Kosem for their expertise. #DataScience #OCR #NLP #Slovenian #MachineLearning #ComputerVision #Benchmark

Prejšnji teden smo v okviru dogodka 𝗦𝗟𝗔𝗜𝗙: 𝗨𝗺𝗲𝘁𝗻𝗮 𝗶𝗻𝘁𝗲𝗹𝗶𝗴𝗲𝗻𝗰𝗮 𝘃 𝗠𝗦𝗣 delili svoje izkušnje o vpeljavi umetne inteligence v podjetja. Andrej Miščič je predstavil svoje razmišljanje o tematiki in delil tudi nekaj praktičnih primerov uspešnih zgodb. Ključno sporočilo? Najprej identificiramo probleme, nato preverimo, ali jih lahko rešimo z umetno inteligenco – in ne obratno! Če tudi v vašem podjetju razmišljate o uvedbi umetne inteligence, se z veseljem pogovorimo o naslednjih korakih. Hvala za vabilo Združenje za informatiko in telekomunikacije, Slovenian AI Factory - SLAIF!

One of our co-founders Andrej Miščič will be speaking at the upcoming Claude Masterclass event organized by AI HUB Slovenija. If you're exploring how AI fits into your workflow, this one's worth your time. 📅 10.6.2026, 13:00-16:00 🔗 https://luma.com/sckjadr1

Big shoutout to our Blaž Pridgar, who shared his expertise on Homography at the DataScience@UL-FRI Spark sessions meetup this Tuesday! He walked attendees through the theory and demonstrated how this technique can be applied in real-world scenarios, including how our team used it to compute the velocity of an object on video for one of our clients. Want the full breakdown? We'll be publishing a blog post with all the details soon! 👀

Maja začnemo z drugo izvedbo Akademije velikih jezikovnih modelov. Več informacij in prijave 👇

📐 𝗛𝗼𝘄 𝗱𝗼 𝘆𝗼𝘂 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗿𝗲𝗮𝗱 𝗮𝗻 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸? In our latest blog post, we dig into the details behind LES (Lestvica embeddingov za slovenščino), starting with a question that's easy to overlook but surprisingly important: 𝘩𝘰𝘸 𝘥𝘰 𝘮𝘰𝘥𝘦𝘭𝘴 𝘵𝘰𝘬𝘦𝘯𝘪𝘻𝘦 𝘚𝘭𝘰𝘷𝘦𝘯𝘪𝘢𝘯 𝘵𝘦𝘹𝘵? Not all tokenizers treat Slovenian equally – and that has real consequences for performance. We also break down why single-language benchmarks matter and how to interpret the metrics in LES, so you can make more informed model choices. 📝 Read the full post: https://lnkd.in/dJwdBY8h 📊 Explore the leaderboard: https://les.valira.ai/ Author: Neli Čatar 💡 If you're working with Slovenian language data and wondering which embedding model to pick – this one's for you. #NLP #Embeddings #Slovenian #MachineLearning #Benchmark #LES #DataScience #Tokenization

Last week, our Data Scientist Maja Kolar gave a talk at a local meetup, where she shared her experience and perspectives on using AI tools for programming. 👉 Key takeaway: AI agents are most effective when you have a clear understanding of the problem you’re solving. If the problem itself isn’t well defined, they may not always lead you to the right solution — even if they sound convincing. But when used with clear intent, AI agents can be a powerful addition to your toolbox and often help you get there faster. Always great to catch up with the local community and share experience. Thank you for the invite, Klub Ada!