Work with measurable results: two in-depth case studies plus the rest of the projects with their code, metrics and links.
A ~2 KB differentiable filter that undoes an illumination shift before the image reaches a frozen detector. No retraining, no labels, minutes on a CPU.
An LLM + RAG architecture that builds a verified ontology incrementally and uses it to flag false claims in films, podcasts and texts. Published at DCAI'25.
Defect detection and segmentation (CNNs, ViT, YOLO, RF-DETR) plus sidewall reading with a LoRA-tuned GLM-OCR, in production at VRAIA. Served with vLLM and Docker.
LoRA adaptation of the multimodal GLM-OCR model (LLaMA-Factory) that returns structured JSON from the image of a receipt or an invoice. Fully local inference.
Training, evaluation and embedding extraction with Meta's ViT backbones: full fine-tuning, linear probe or raw feature extractor, with per-class metrics and checkpoints.
Thousands of HYSPLIT back-trajectories on ERA5 and DBSCAN clustering to locate blockings from the actual stagnation of air, without Eulerian thresholds. Validated on the 2003 and 2019 heatwaves.
A fully local console agent with native tools (files, web search, Python execution) and RAG over your own documents with ChromaDB.