João Cardeira

I am a PhD student at NOVA LINCS, NOVA School of Science and Technology, working on large vision-language models, video and anomaly understanding, and multilingual LLMs.

My research moves video anomaly analysis beyond detection, toward models that can explain what is anomalous and why (FineVAU, AAAI 2026), and explores how LVLMs can use their own self-reflection to handle long videos efficiently (Self-ReS, ICME 2025).

I am also part of the core development team of AMALIA, the fully open large language model built for the Portuguese government. I contributed to AMALIA-VL — the first open-source vision-language model built natively for European Portuguese — working on the training pipeline, pt-PT multimodal data mix, and evaluation suite, and I am first author of PorTEXTO, a European Portuguese benchmark for visual text extraction.

More broadly, my research interests are centered on video and anomaly understanding — building systems that can watch, interpret, and reason about what happens in video, especially when things go wrong. I’m drawn to problems where vision and language meet: grounding explanations in visual evidence, handling long temporal contexts, and evaluating models on nuanced, real-world scenarios.

I care about open, sovereign AI: models, data, and benchmarks that anyone can inspect and build on, especially for languages and variants that mainstream models underserve.

Feel free to reach out — I am always happy to talk about multimodal models, pt-PT NLP, or video understanding.