What Are Reasoning LLMs?
Chain-of-thought traces, reinforcement learning from verifiable rewards, inference-time scaling — the technology behind…
Archive
Blog
Notes on research, methods, papers, and technical work.
Type to search across blog posts.
Fine-tuning small models for drug safety: when, why, and how much data you actually need
In pharmacovigilance, a well-adapted small model often outperforms a much larger general-purpose one. Practical guidanc…
On Detecting AI Writing
Perplexity scores, watermark tokens, classifier confidence — the technology behind AI text detection is fascinating and…
Multi-agent LLM architectures for drug safety workflows
Monolithic prompting breaks down in pharmacovigilance. Multi-agent orchestration — where specialized LLM agents handle …
How LLMs handle (and mishandle) medical terminology and ontologies
LLMs can process medical language impressively, but their relationship with controlled vocabularies like MedDRA is more…
Can LLMs detect negation reliably in safety narratives? A practical look
Negation is one of the hardest linguistic problems in pharmacovigilance NLP. LLMs improve on older methods but still fa…
Context windows, chunking strategies, and what gets lost when safety documents are too long
Drug labels, long case series, and multi-page narratives routinely exceed practical LLM context limits. How you chunk t…
Stimulated reporting, Weber effects, and what happens to disproportionality after media coverage
Media coverage, regulatory warnings, and the passage of time all distort spontaneous reporting data. Understanding thes…
Pediatric pharmacovigilance: why children are not small adults, statistically or clinically
Children present unique challenges for drug safety surveillance. From developmental pharmacokinetics to off-label presc…
Benchmarking signal detection algorithms: the reference set problem
Evaluating pharmacovigilance signal detection methods requires a reference set of known positive and negative drug–even…
Postmarketing Spontaneous Reporting Systems: the Backbone of Real-World Drug Safety Surveillance
A practical overview of postmarketing spontaneous reporting systems, how they work, why they matter in pharmacovigilanc…
How a diabetes drug affects the brain
GLP-1 drugs such as semaglutide and liraglutide are changing how we think about diabetes treatment, appetite, reward, a…
Artificial intelligence and large language models in drug safety
How AI and large language models are helping turn drug labeling and other safety text into faster, more usable pharmaco…
Why retrieval-augmented generation matters more than model size in pharmacovigilance
In drug safety, grounding LLM outputs in authoritative sources is more important than scaling model parameters. RAG off…
Evaluating LLMs for pharmacovigilance: beyond F1 scores
Aggregate metrics like F1 can hide exactly the failures that matter most in drug safety. Better evaluation means measur…
When drug safety LLMs stop being demos and start becoming infrastructure
A blog post on what three recent papers reveal about the real future of LLMs in pharmacovigilance and medication safety.
Signal detection in pharmacovigilance
A short introduction to PRR, ROR, IC, and MGPS in pharmacovigilance, and why signal detection metrics matter in spontan…