A tiered curriculum
Foundation through Practitioner and specialist tracks, across every clinical cadre — not a single course for doctors only.
// ki-cai
Kenya Institute for Clinical AI — a proposal for the institution that teaches Kenya's doctors, surgeons and nurses to use clinical AI, and that measures whether the teaching worked.
$ cat ask.txt
Approve Phase 0 — USD 1.94 m over months 0–12. Nine founding posts, the
simulation centre and sandbox build, and negotiation of competency standards with the two
professional councils. It is the only decision that commits capital.
// the_case
In 2025, a study across fifteen primary-care clinics in Nairobi examined nearly 40,000 patient visits in which clinicians had AI decision support in the electronic record. It reported a 16% relative reduction in diagnostic errors and a 13% relative reduction in treatment errors among clinicians who engaged with the tool. That signal was generated here, on our case mix, by our clinicians.
Read the same work carefully and a second finding sits alongside the first. The benefit was concentrated among the clinicians who used the tool well. The technology did not distribute its benefit evenly — it distributed it to the people who knew how to use it.
The variable that determines whether clinical AI helps or harms a Kenyan
patient is not the model. It is the training of the clinician sitting in front of it.
And the obvious response is not sufficient. A randomised trial published in 2025 took physicians who had already completed twenty hours of AI-literacy training, gave them clinical cases, and exposed half of them to deliberately erroneous model output. They deferred to it. Prior literacy training did not protect them: confidence in the tool had been raised, scepticism had not. Teaching clinicians what AI can do, without rigorously training the reflex of doubt, produces a clinician more dangerous than the one you started with.
// what_is_being_built
An assessment that can fail people
An AI-OSCE built around one hidden error, with the pass mark set by a modified Angoff panel rather than an arbitrary 50%.
A sandbox and simulation centre
Where clinicians meet erroneous model output under supervision, before they meet it on a ward round.
An evaluation programme
Kirkpatrick levels 3 and 4 — behaviour and patient outcome — published whatever they find, or we admit we do not know.
// the_numbers
Five-year model. Every figure computed from a published assumption register, not typed into a table by hand.
| Five-year expenditure | USD 17.37 m |
| Five-year earned income | USD 5.34 m |
| Five-year subsidy required | USD 12.03 m |
| Learners certificated | 21,746 |
| Blended cost per learner | USD 799 |
| Subsidy per learner | USD 554 |
| Earned-income cover, Year 5 | 49% |
| Core establishment at steady state | 71 FTE |
The honest headline: this does not wash its own face, and it is not designed
to.
At fee levels Kenyan clinicians can actually pay, earned income reaches 49% of cost by Year 5. Full recovery would need 2.1× the modelled income. The Institute is a subsidised public good with a substantial and growing earned-income component — not a business. That is the argument to accept or reject. Everything else is detail.
At fee levels Kenyan clinicians can actually pay, earned income reaches 49% of cost by Year 5. Full recovery would need 2.1× the modelled income. The Institute is a subsidised public good with a substantial and growing earned-income component — not a business. That is the argument to accept or reject. Everything else is detail.
// the_documents
Blueprint 45 pp
Annex A — Business case
Annex B — Competency standards
Annex C — Level 1 common core
Facilitator deck
// the_argument_in_order
The blueprint is the proposal. These five essays are the reasoning underneath it — read the first, then whichever of the four you most want to argue with.
Another Arrow in the Quiver: The Clinical AI Institute Kenya Should Build
The proposal itself: the case, the institution, the curriculum, the sequence, and what could go wrong. Start here.
Read →Borrowed From an Art School: Where the Clinical AI Framework Actually Came From
Where the competency framework came from, and why its provenance outside medicine is the reason it transfers to clinical work.
Read →One Hidden Error: What an OSCE Is, and What an AI-OSCE Would Be
What an OSCE is, and what an AI-OSCE would be — the examination the whole blueprint rests on.
Read →The Angoff Panel for Testing Clinicians
Where the pass mark comes from, and why an arbitrary 50% cannot be defended when a certificate is a safety claim.
Read →Measuring What Actually Matters: Kirkpatrick 3 and 4 for Student-Clinicians
Kirkpatrick levels 3 and 4: how you find out whether any of it changed behaviour or helped a patient.
Read →// who_is_proposing_this
Dr Neal Aggarwal is a physician and data scientist based in Nairobi. He has trained more than 600 AI practitioners — from medical doctors to school leavers — to build and deploy real machine learning systems, and founded CyberCollege, the first online college in Africa, pioneering the flipped classroom there before the word MOOC existed. His work applies machine learning to medical diagnostics, clinical NLP and health-system process engineering.
The clinical tools on this site are built, not sketched: DR Detector performs offline diabetic-retinopathy screening on a smartphone with no connectivity and no patient data leaving the device. That is the standard the Institute's teaching is meant to hold.
This proposal is written in a personal capacity. It is published in full — including the parts that are uncomfortable — because a proposal that cannot survive being read is not worth funding.
Interested institutionally?
Host institutions, professional councils, funders and ministries — I would rather this were argued with than admired.
Start a conversation →