I ship production LLM systems
in healthcare.
Research Software Engineer specializing in clinical NLP. 3-week concept-to-production cycles. Pipeline speedups measured in orders of magnitude. Code running in clinical trials.
120 days → 4 hours
100% schema validity
now in clinical trial
via synthetic data
Featured Work
CRANE-style Structured Medical JSON
Built constrained decoding pipeline converting free-text pathology reports to schema-valid CAPeCC JSON. Implemented free-reason → switch-token → JSON window generation pattern. Diagnosed and patched Outlines → llama.cpp logits incompatibility breaking FSM-based token masking. Ported from Hugging Face to llama.cpp (GGUF) for large-model inference on A100s.
Caring Contacts — Psychiatric Discharge
Collaborated with Psychiatry to build RAG system generating personalized post-discharge hope letters from patient charts. Proposition-level chunking + two-stage retrieval (BM25 → Qwen reranker) with clinician-defined safety guardrails. Shipped concept to production in 3 weeks.
About
I'm a Research Software Engineer at Sunnybrook Research Institute in Toronto, where I'm the sole NLP specialist building production AI systems for clinical applications.
My work sits at the intersection of LLMs and healthcare: constrained decoding for structured report generation, RAG systems with clinical safety guardrails, and large-scale PHI de-identification pipelines.
Background in biomedical engineering. I learn by building — typically implementing research papers within days of reading them. I optimize for asymmetric upside and ship fast to learn fast.
Let's talk
Looking for Applied AI / Research Engineer roles at top-tier labs.