AI·ML Engineer
medical imaging · DICOM · PACSI build production-grade medical-imaging AI — closing the gap between research and the radiology reading room.
Signal
6
production ML systems in clinical use
0.9983
F1 — mammography laterality classifier
~75s
per spine study · dual Tesla P40
122K
German medical terms · live dictation corrector
Selected work
aycan Medical Systems / Paratus · 2025
Automated vertebral and disc segmentation that renders labels straight into the radiologist's PACS viewer — C1 to sacrum, in minutes, demoed at RSNA 2025.
aycan Medical Systems / Paratus · 2025
A vision-language model that finds and redacts burned-in PHI from DICOM pixel data on-premise — a novel application of Florence-2 with no published prior art.
aycan Medical Systems / Paratus · 2025
Real-time radiology dictation with a from-scratch German medical text corrector — dual-Whisper transcription, structured LLM reports, fully local inference.
aycan Medical Systems / Paratus · 2025
A three-model stacking ensemble that recovers laterality and view position from real-world mammography — trust-gated to lean on metadata when it's reliable and pixels when it isn't.
Live signal
Public output, pulled at build time — no curation. ORCID, GitHub, and Hugging Face, rebuilt nightly.
GitHub · top repositories
Hugging Face · 2 public Spaces — arXivGPT · SegmentAnythingxGroundingDINO
Experiments
A custom ShinkaEvolve fork where a frontier-LLM orchestrator (Claude Opus) directs the search and spawns evolutionary child solutions as containerized GLM-5 agents. Pointed at open targets — faster llama.cpp inference, the 4×4 matrix-multiplication record (AlphaEvolve's 48 multiplies), Erdős problems, and bitpacking. Research-stage: the matmul search space proved too large to beat the record, but the distributed agent-evolution harness is the payoff.
ShinkaEvolve · LLM orchestrator · GLM-5 agents · AlphaEvolve · llama.cpp · research
Open-vocabulary detection and segmentation in the browser — upload an image, name what you're after, and get masks and boxes back. Runs as a live Hugging Face Space.
SAM · GroundingDINO · open-vocab · Gradio
A DeepSeek-OCR-inspired optical encoder that bridges a frozen DeepEncoder to Qwen3-VL through a universal MLP adapter — compressing pages of text into vision tokens.
DeepSeek-OCR · Qwen3-VL · adapter · PyTorch