Work

Projects are indexed by the question that caused them to exist. Each record describes the constraint, implementation, evidence, and current limits.

Computer Vision

• Can ordinary video detect infantile epileptic spasms? — action recognition and optical flow on long clinical video.
• Can action recognition run entirely in a browser? — X3D-M inference through ONNX Runtime Web.
• Can classical vision detect document tables in under 100 ms? — morphology, whitespace, and hybrid detection.
• How much segmentation can be pruned before structure is lost? — accuracy, mIoU, and throughput measured together.
• Can monocular depth add useful distance to object detection? — YOLOv12 combined with MiDaS depth estimates.
• Can a robot predict future visual states in latent space? — latent dynamics and visual reconstruction.

Runtime Systems

• Can Echo State Networks become practical at the edge? — C++17, custom CUDA kernels, and sparse reservoirs.
• Can model loading overlap inference? — a measured negative result.
• Can vLLM have a smaller operational interface? — single-machine serving works; multi-node orchestration is unfinished.
• Can test-time computation extend a frozen language model? — architecture and infrastructure exist; training remains.

Applied Machine Learning

• Can speaker-aware transcription remain entirely offline? — local Whisper and PyAnnote inference.
• Can BERT adaptation and serving fit a smaller training budget? — LoRA training and a production serving path.