Six years turning research into things that actually run. The constant across all of it: pixels, and everything you can bolt onto them. Classic computer vision, then edge deployment on Jetson / TensorRT, then MLOps that survives production, and lately LLM / RAG / VLM / agentic systems — same throughline, I just kept finding new modalities to feed the models. One peer-reviewed paper I'm genuinely happy about (HiPEAC 2025, on active inference), tech-lead experience on an EU edge-AI project, and a standing suspicion of anything a vendor describes as "just."

I like problems where the data is rich and slightly unruly — images, video, text-and-vision together. The interesting failures live there: a model that's fluent in one modality and clueless the moment you cross into another. Most of applied ML is really about those seams, and I've spent my career happily poking at them.

• Vision & multimodal — from classic CV pipelines to VLMs: detection, recognition, retrieval, and getting language and images to actually agree on what they're looking at.
• From lab to the real world — models compressed and quantized until they fit a robot's power budget, then deployed where latency has consequences. The gap between a benchmark and a deployment is where the real work hides.
• Systems over demos — the interesting part of an ML system is never the model. It's the seam where it meets a messy world and has to pretend it's coping.
• Explaining hard things — interactive explainers at makesnosense.io: positional encodings you can poke, Kalman filters that stop being scary. Teaching is just debugging your own understanding in public.

_{A chatbot with no model, no server, and no ambition. Click a question.}