Writing again

For the last decade I have built AI platforms behind NDAs and spoken about the outcomes at conferences, which is a polite way of saying I have let a lot of interesting lessons evaporate between the work and the audience. That is the gap I am trying to close here.

The part I care about is the stretch most decks skip past. Most of what I have shipped was built there: knowledge graphs over literature that looked clean on a slide and fought back the moment a domain expert asked a sharp question; patient digital twins that looked nothing like real patients until the conditioning was reworked from the disease up; portfolio intelligence systems that had to survive a BD team's scrutiny at 9am on a Monday with no caveats to hide behind.

What you'll see here

• How agentic workflows actually earn a clinician's trust — the design choices that turn an impressive demo into a tool someone is still using in six months.
• Where digital twins help, and where they mislead with confidence — grounded in prototypes I have built, not in slides I have seen.
• Notes from building QuriousRI and AlphaMeld — the parts I wish someone had written down before I learned them the hard way.
• Talks I am preparing, papers I am reading, decisions I am still second-guessing — thinking in public rather than waiting for certainty.

The house style

Short posts. No listicles. Opinionated, citable, and willing to be wrong in writing. Diagrams when they earn their place.

If any of this is useful, get in touch.