Does the AI hold up where the medicine actually happens?
From May to August 2026, I worked as a medical domain expert at SME Careers, evaluating medical LLM outputs for clinical accuracy, safety and guideline alignment. I identified hallucinations, reasoning failures and unsafe recommendations, and wrote rubric-based feedback for evaluation and red-teaming.
The samples below are independent portfolio work.
Portfolio samples
Deploy Check — Health-AI Deployment Scenarios
Eight health-AI products pitched for clinics with unreliable power and no specialist on site. You make the call, then see what breaks.
Testing a general model on clinical safety
Drug dosing, red-flag symptoms, crisis response and vague-symptom triage, run against a general-purpose model and scored on a fixed rubric.
A supplement article, checked line by line
A published brand article reviewed for clinical accuracy. Three high-risk interaction omissions found, with corrected language written for each.
Energy supplement — before and after
The original AI-generated article, the clinical and regulatory-risk findings, and the corrected version side by side. Brand name redacted.
AI-generated health content — a review case study
The original draft, the issues identified, and the edited version with a summary of every correction.
Falsely flagged by an AI detector?
How detectors actually reach a verdict, the published error rates that undermine them, and how to open an informed conversation with a department. Free.
How an evaluation is actually done
The method is the product. A transparent evaluation is more persuasive than a large claim about AI safety.
What I test
Dosing, red-flag symptoms, crisis response, contraindications, vague-symptom triage and psychiatric frameworks — the tasks where a wrong answer is not a rounding error.
How I score
A fixed rubric applied to every response, with the model and version recorded. Findings are reported exactly as measured.
What I will not claim
A sample audit is not a certification. I report what a test can and cannot establish, and I do not imply a product is clinically safe unless the method supports it.
Essays on this subject
The arguments behind the method — what clinical AI gets right, where it fails, and why the failures matter. Each piece lives on the blog.
How I Made ToluWhispr
I couldn't find a dictation app that was fast, cheap, and worked globally on my PC. So I built one with AI — and I don't know how to code.
Feels Like Writing Is Dead
A machine wrote a sentence I could not have written, and it was beautiful. On what happens to the value of writing when competent prose stops being scarce.
The Wrong Object
Psychiatric evidence isn't found in the patient. It's made in the room, which may explain why AI has reached every diagnostic frontier except this one.