PhD candidate building RAG, imaging, and on-premise LLM systems that fit into real clinical workflows — not around them.
Model-in-the-loop systems that fit existing clinical workflows instead of replacing them.
Real-time risk stratification for acute care.
Computer vision for MRI, CT, and echo video.
Robustness and interpretability across messy, heterogeneous hospital data.
A self-hosted RAG chatbot for clinical nutrition counselling — swappable local or cloud models, and topic-boosted retrieval that actually finds the right passage.
There's a paused side-experiment in here too: a lightweight I-JEPA vision encoder for scanning food photos, shelved once the retrieval work proved more valuable to ship first.
An on-premise model that turns clinical notes into structured ADIME documentation — no cloud, no per-token cost, and no PHI leaving the hospital network.
Running on Qwen 14B with Pydantic + Instructor for structured extraction, validated against 41 golden clinical cases before anything touches a real chart.
A deterministic cardiovascular risk score paired with an echo-video MI classifier and an LLM that drafts the clinician report.
Every AI verdict ships next to the numbers it's grounded in — the point isn't a black-box score, it's a report a clinician can actually check.
A personal server for training, local inference, and self-hosting — tunnelled to the internet with zero open ports.
Adapting Qwen2 to nutrition tasks with 4-bit QLoRA via Unsloth, exported straight to GGUF for local Ollama serving.
Comparing CNN, LSTM, XGBoost, and optical-flow approaches on EchoNet-Dynamic and HMC-QU for cardiac motion classification.
Serving fine-tuned, quantized models fully offline on consumer GPU hardware via Ollama and llama.cpp.
Open to healthcare-AI research collaborations, clinical deployments, and speaking engagements.