ML pipelines and applications.

I designed the coverage-table extraction pipeline and PDF review interface. Rewriting the application core with Rust and Python reduced RAM usage by roughly 70%.

I adapted BiRefNet to client images and built the annotation-to-inference pipeline. Visually observed defects: roughly 50% before, below 25% after.

I am building a visual editor with reusable primitives, a timeline, previews and video rendering to compose Math / ML animations without coding every scene from scratch.

I built a data preparation component for R&D: HDF5 reading, trajectory materialization in Parquet and MP4 video export.

I compared models using V-JEPA embeddings with temporal baselines. Time improves some tasks but can also explain scores without visual error detection.

I prepared data and experimented with LeWM training on a DROID 10k subset. The experiment did not achieve the robot-arm control objective.

I developed the equestrian management application from frontend to backend and integrated YOLO to explore rider pose. LLM recommendations remain experimental.