Model training
Bringing new architectures into Roboflow training as they come out: YOLOv8 instance segmentation, YOLO11, YOLOv12, YOLO26, and RF-DETR from the first release through the larger sizes and segmentation.
Real-time detection transformer on a DINOv2 backbone. Instead of hand-picking an architecture, fine-tune once and evaluate thousands of speed/accuracy trade-offs without retraining. Does segmentation too.
100 detection datasets of things VLMs rarely saw in pre-training. Strong zero-shot models get under 2% on the hard medical sets. The goal: teach a model a new concept from a few examples and written instructions.
Bringing new architectures into Roboflow training as they come out: YOLOv8 instance segmentation, YOLO11, YOLOv12, YOLO26, and RF-DETR from the first release through the larger sizes and segmentation.
PaliGemma 2 and Florence-2 fine-tuning with LoRA, plus the training and deployment path behind it. Cosmos 3 Edge after that, on images and video.
The parts every job shares: early stopping, automatic batch sizing for RF-DETR, out-of-memory fixes, and training more than one model on the same dataset version.
Built the research team’s internal experiment platform. Launch training runs on our GPU cluster from a CLI, fan a benchmark like Roboflow100-VL out to a hundred jobs, and keep results and artifacts in one place.
PatchCore and FoundAD. Only needs images of good parts, and gives you heatmaps.
Train action recognition on your own video in Roboflow. Label time ranges, fine-tune Cosmos 3 Edge or V-JEPA 2.1, then deploy to the cloud or your own GPUs.
The search from the RF-DETR paper. Run it on your own dataset, for your own hardware.
Roboflow Inference, the open-source server. I wrote the async model manager and added serving for PaliGemma, Florence-2, SAM 2, OWLv2, Perception Encoder, a lot of YOLOs, RF-DETR and the anomaly models.
Worked on the SAM 2 backend for Smart Polygon in the labeling tool, including multi-polygon masks. Stood up the SAM 3 video tracker behind click-to-track video annotation.
Contributing to Nebula, the contributor platform behind Roboflow Labs. Building inference infrastructure for AI-assisted labeling and leading a data project.
Building a team of incredibly cracked MLEs and researchers.
Part of Roboflow’s partnership work with NVIDIA, through the Cosmos Coalition, and with other companies. Brought Cosmos 3 post-training to Roboflow.
Two fixes that unblocked PaliGemma fine-tuning. The attention mask let the model see its own labels during training; the suffix is now causal. On PaliGemma 2, the processor dropped the end-of-sequence token from the suffix when an image token was present.
Scaled-dot-product attention fallback.
Cryptographically secure random IDs for new documents.