writing

Jan 2026

Do LLMs Think in English?

Models use multilingual concepts as rhetorical devices, not reasoning primitives.

Oct 2025

Building a ML Pipeline for Microcalcification Classification in OMOP

OMOP's non-auto-incrementing PKs taught me more about database design than any tutorial.

Sep 2025

On Building Augmented Datasets: A Practical Case Study

Why hospital-specific training data matters more than MIMIC.

Jul 2025

When 75% Isn't Enough: Trying to Distill GPT-4 into GLiNER

Prompted probabilities aren't real probabilities. That's why this failed.

Jun 2025

Road to a SOTA PII Model

From PhysioNet heuristics tagging 'pain' as a name to 75% F1.

Jun 2025

Using Decoder-Only LLMs for PHI De-Identification: A Minimal Setup

Skip asking for character positions—just get spans and regex the rest.

Jun 2025

RAG Experiments: Chunking, Retrieval, Reformulation

Proposition-level chunking beat paragraphs for pulling personal details from clinical notes.

Jan 2025

Framing Survival Prediction as Next-Token: A Failed Experiment

Cell2Sentence orders genes by expression. Could clinical features be ordered the same way?

Jan 2025

Building a Bookkeeping Agent: What Worked, What Didn't

I built tools for a financial agent from scratch while interviewing for an agents role. Calculator failed, DuckDB saved the day, and I found my $70/month Starbucks habit.

Dec 2024

MedColBERT: Late Interaction Retrieval for Clinical Text

JaColBERT's training recipe fixed ColBERT's in-domain underperformance. Could it work for medical retrieval?