Built and maintained the full suite of annotation data services — covering import/export across 10+ formats, video annotation tooling, VQA pipelines, and augmentation preview. These services sit at the core of the Datature platform, enabling seamless data exchange between Datature and every major labelling tool.
Highlights
▸Annotation import/export for 10+ formats: COCO, LabelMe, SuperAnnotate, ScaleAI, OpenAI, VoTT, Supervisely, IBM — with multipolygon, sliding window, and 3D NIfTI support
▸VQA/Visual Intelligence annotation pipeline: JSONL and ViFull import/export, phrase grounding with caption/bbox support, Neural Insights
▸Video interpolation service — propagates annotations across intermediate frames
▸SAM2-based video tracking service with PubSub queue and multiprocessing to prevent GPU memory leaks
▸Mosaic and CutMix augmentation preview for training data quality assurance
▸3D NIfTI annotation import/export: class-indexed and one-hot mask exports, 3D TFRecord with bitmask3d
▸KEDA scale-from-zero for video services — optimises cost during low-traffic periods