writing

Agent-based modelling gets genuinely complex behaviour out of agents following hard-coded rules — no neural network anywhere. Builds a highway traffic simulation in Mesa where a binary move-or-wait decision per car is enough to produce phantom jams and gridlock on its own, and looks at where that still beats a learned model.

A clinical trial protocol is a hundred pages of prose that every downstream system has to re-read by hand. Walks through a pipeline that turns one into CDISC USDM v4 — layout-aware parsing, LLM section mapping against curated dictionaries, then schema graph linking — with an expert review gate in the middle. Extraction drops from about 40 hours to 2.

A 3B model beating an 8B one isn't a story about model size, it's a story about prompting. Programs Llama-3.2-3B with DSPy signatures, chain-of-thought and automatic few-shot selection to reach 58.57% on astronautics Q&A — against 55% for an unoptimised 8B, at a fraction of the compute.

Clinical data can't leave the building, which is exactly what makes it hard to fine-tune a model on. Proposes MASDO, a four-agent pipeline that generates, manipulates, labels and validates tabular clinical records between them — so a team can train on realistic data without a real patient record ever being involved.