Thoughts on
data science.
Reflections & project notes
Where Regex Ends: Practical NER for PII in Free Text
Running spaCy NER (via Presidio) to catch names in free text — model selection as a recall dial, the confidence score you can't threshold, domain shift, and why de-identification is a recall-first problem.
↗
한 시간 만에 만든 한국어 AI 말투 탐지기 (그리고 그 탐지기가 이 글을 채점한 결과)
내 블로그 글에서 AI 냄새가 나면 슬랙으로 알려주는 파이프라인을 하루 만에 만들었다. 첫 실험 대상은 사람이 쓴 내 글, 두 번째는 AI와 함께 쓴 이 글이다.
↗
NoteGuard: a One-Month Retro
Building the trust layer for clinical AI — from a hackathon hack to a deployed, honestly-measured de-identification agent.
↗
From Build-in-a-Day to v1.1.0: Two Months of NHS Policy Navigator
A retro on the two-month journey of NHS Policy Navigator — from a one-day MongoDB hackathon build in May, through right-sizing the stack for a public-data PoC, to a Gold RAP product with a Query Router (v1.1.0) presented as a prototype in July.
↗
Candy, Sweets, and Jelly: Explaining Data Science at School (and Getting My Labels Wrong)
I volunteered at two school events to explain what a data scientist does — to Year 1 and 2 at an academy, and to Year 7–8 and Year 12–13 at a secondary school. Part of my talk was a sweet-sorting classifier, and it accidentally demonstrated why label vocabulary matters more than model choice.
↗
Survey data analysis for free-text responses
24년 초반으로 기억하고 있다. 커밋 히스토리는 그 해 가을이더라.
↗
Four Knobs in My Text Preprocessing Pipeline (and What Each One Actually Does)
In What I Got Wrong Building a Sentiment Analysis Pipeline for Survey Data and the follow-up checklist, I w...
↗
What I Got Wrong Building a Sentiment Analysis Pipeline for Survey Data
I ran a sentiment analysis pipeline over free-text responses from an internal survey asking people for thei...
↗
Fixing a Sentiment Analysis Pipeline: A Checklist for Next Time
In What I Got Wrong Building a Sentiment Analysis Pipeline for Survey Data, I traced why a survey sentiment...
↗
Prompting Gemini to generate Makaton-style symbol cards
What I learnt about prompt engineering for consistent, licensing-safe AAC symbol generation while building the Getting Started with Makaton choice board.
↗
On LinkedIn