Writing
Essays on AI-oncology, regulatory strategy, and the discipline behind calibrated judgment.
Each piece below is written under the Citation-Required Output Contract — every claim of consequence is sourced, and the reasoning trail is shipped alongside the conclusion. The standard is the same one applied to client deliverables. Most of the regulatory detail here is FDA, because most published thinking is; where the answer changes between FDA, EMA, PMDA and NMPA, the essays say so rather than generalising from one agency.
- 2026-09-23 · On AI in trial design
Read the Limitations Section: What Three AI Trial-Design Papers Say About Their Own Limits
TrialAtlas, an Oxford offline-policy benchmark and a Stanford virtual biotech in Science all published inside two months. All three are genuinely useful, and all three state their own limit in print: they retrieve and organise precedent well, and do not yet know when precedent is the wrong guide. The failure that costs most is not a bad answer — it is a team that deploys on the headline, fails on the caveat, and concludes AI does not work for drug development.
- 2026-09-07 · On trial geography
Where a First-in-Human Trial Begins Is Also a Decision That Is Not Yours
The debate about early-phase research leaving the United States is conducted in one currency: how fast and how cheaply a trial can start. That is the second question. The first is evidentiary — whose prior questions is the data being built to answer. The geography decision moves the patient's chair, the developer's dollar and the investor's tranche at once, and not in the same direction; ORIENT-11 shows what it costs when the room a trial was designed for is not the room that adjudicates it.
- 2026-09-04 · For boards and investors
Six Questions a Board Should Be Able to Answer Before Approving the Next Cohort
The six gates the FIH Diligence Read runs on a first-in-human oncology programme, published in full: whether the mechanism is specified tightly enough to be wrong, whether exposure was ever characterised, what the dose was compared against, whether the signal would survive a control arm, whether the safety database is old enough to mean anything, and whether the next experiment resolves the uncertainty or merely advances the programme. Each closes on what a failure does to the commitment made at the last raise.
- 2026-09-02 · Regulatory note
The Decision Is Not Yours — and It Is Not One Decision
ICH M15 and the joint FDA/EMA principles have genuinely harmonised how a model's credibility is judged. But credibility depends on context of use, and context of use is a property of your programme rather than of the framework — so dose, population and endpoint are still adjudicated regionally. One sintilimab dataset, approved in China and sent back by an FDA advisory committee, shows what that costs a programme that finds out late.
- 2026-09-01 · Regulatory note
What a First-in-Human Trial Is Actually For
An FIH oncology trial is worth exactly the decisions it lets three people resolve: the patient who said yes, the person spending the next development dollar, and the person spending the next investment dollar. Why the escalation algorithm is subordinate to the decision sequence; why dose optimization and the June 2026 QSP/MABEL draft moved the standard the installed base was designed against; and why the divergence between those chairs is the finding, not the noise.
- 2026-08-24 · Regulatory note
Model Validation, Process Validation, and the Decision That Is Not Yours to Make
Process validation asks whether a model improves a consequential decision — which holds while the decision is yours. At the regulatory interface it isn't: the agency decides whether the evidence supports the claim. The same model output can occupy four regulatory positions whose evidentiary burdens differ by more than an order of magnitude, and a program can slide between them without anyone deciding it should.
- 2026-08-06 · Regulatory note
What an Oncology Pre-IND Meeting Actually Decides — and How to Prepare for Useful FDA Feedback
The PDUFA VII sequence from readiness to Request for Clarification; why the meeting package does most of the substantive work; how to spend a ten-question ceiling; what the division assesses, in a common core plus conditional oncology modules; and why pre-IND advice is not permission to dose patients.
- 2026-08-05 · Engagement note
Why and When Should an AI-Driven Biotech Hire a Fractional CMO?
The gap a fractional CMO exists to close for an AI-driven biotech, the signals that say hire one now, the honest cases where it is the wrong call, and the AI-platform variant. What the role is — and the three things it is not.
- 2026-08-02 · Methodology case study
What a Governed Multi-Model Workflow Caught in Its Own Oncology Analysis
A bounded case study of role separation, frozen evidence, adversarial cross-family verification, deterministic controls, and human gates — including what one uncontrolled pilot can and cannot establish.
- 2026-05-12 · Thesis
Three Ways AI-Oncology Companies Will Fail at the FDA — and the Clinical Expertise Gap That Causes It
Within twelve to eighteen months, a meaningful share of today's well-funded AI-oncology companies will hit the FDA wall. Three failure modes — the endpoint that never was a drug, the patient who was never going to enroll, the black box that cannot be audited — all trace to the same missing input: calibration. (Also on LinkedIn.)
- 2026-05-20 · Follow-up
How I Would Know If I'm Wrong — Six Conditions for Falsifying the Three Ways Thesis
The six conditions under which the Three Ways thesis would be wrong — volume, distribution, causation, closure, timing, and calibrated-departure traction. Each tested against a specific part of the thesis. A calendar-anchored commitment to publish two post-mortems — at twelve and twenty-four months — whether the thesis holds or fails. (Also on LinkedIn.)
- 2026-05-12 · Engagement note
What the First 90 Days Look Like — Biotech-Shaped AI-Oncology Companies
The canonical four-phase, twelve-week operating arc for biotech-shaped AI-oncology companies — those running their own programs toward an IND. Each phase produces one specific artifact and retires one specific decision. The arc is built on an expansive-then-decisive discipline inside Phase 2 to guard against premature lock-in.
- 2026-05-12 · Engagement note
What the First 90 Days Look Like — AI-Platform Companies Selling Into Pharma
The parallel arc for AI-platform companies whose commercial trajectory depends on partners successfully defending the platform's role to the FDA. The agency-facing work in the biotech variant becomes partner-facing work: Platform Validation Package, Partner Q&A Library, Reasoned-Departure Support.
- 2026-05-15 · Companion artifact
AI-Oncology Failure Watch — Locked Seed List
The universe against which the failure thesis will be scored, locked before the outcomes are known: four inclusion criteria, the companies that met all four, the exclusions and why each was made, and the threshold at which the thesis is wrong. Published so the scoring cannot be adjusted after the fact.
For an engagement conversation, see how to start a conversation.