Running a Study Without Writing Code
Ten hospital exports, one AI agent, and the part that takes longest — working out whether we can believe it.
Hsieh-Ting Lin, MD 林協霆
Department of Medical Oncology 腫瘤內科部
Koo Foundation Sun Yat-Sen Cancer Center 和信治癌中心醫院
Department of Medical Oncology 腫瘤內科部
Koo Foundation Sun Yat-Sen Cancer Center 和信治癌中心醫院
The recording
The walkthrough
Data, cleaning, exploring, analysis, writing — five stages
in one continuous terminal session. Twenty-two chapters. This is the one
that was presented.
The long version
The same study with the arguments kept in: what happens when
a prompt is ambiguous, 120 analytic paths, four independent reviewers.
Both play in the browser with no plugin. The text is real text — select any
line and enlarge it. Speed, chapters and fullscreen are in the bar at the
top.
Slides
Deck
What is presented around the recording.
Reference edition
The full mission-by-mission walkthrough behind the talk.
Backup deck
The slides with a still frame from the recording after each
signpost, for presenting without a browser.
The data are entirely synthetic. They are simulated from the
published summaries of a phase 3 trial. No real patient is represented, and
nothing here is evidence about atezolizumab, bevacizumab or sorafenib.