Clinical AI/built here, for here.
Lorraine is built by South African doctors for South African doctors. We started from a frustration we lived — board-exam prep that didn’t fit the CMSA blueprint — and we ended up building a clinical reasoning model calibrated to the healthcare system we actually practise in.
One team,
one context, one model.
- /founded-byPassionate South Africans
- /built-forCMSA trainees, clinicians, clinics.
- /grounded-inSA STG · EML · IMCI · VTP · BANC+
- /modelLor-1, served on our own infrastructure.
- Founded by
- SA registrars
- Clinical model
- Lor-1in production
- Evaluation
- LorBenchSA clinical
- Launched
- March 2026
How we work
Six principles we hold to.
These aren’t values we decided at an offsite. They’re the decisions we keep making in build reviews, in data-quality calls, and in the release-readiness rubric before a checkpoint ships.
- 01
Doctor-first
Every feature is sanity-checked by practising South African clinicians before it ships. If a registrar would not trust it at 3am, it does not ship.
- 02
Local by default
SA drugs, SA formulary, SA guidelines, SA referral ladder. Global context where it helps, local context where it counts.
- 03
Evidence, not vibes
Everything we claim about the model is measured on a benchmark we built for this context. Nothing ships on loss alone — every release is gated by a lift we can point to.
- 04
Honest abstention
If the guidance is ambiguous, the model says so. We would rather surface a gap than produce a confidently wrong answer that nobody can spot.
- 05
Built with the community
Registrars, consultants, and clinical advisors sit inside our release cycle — not in a testimonial at the end of it.
- 06
Always improving
Continuous updates against exam trends, new guidelines, and real clinician feedback. The model in production today is not the one that shipped on day one.
06 principles · updated release over release
Our story
From registrar frustration
to a model for SA practice.
Lorraine started as a tool we wanted for ourselves and grew into clinical infrastructure for a healthcare system that was underserved by foreign-built AI.
- 01 / The gap
Boards prep was scarce, expensive, and not built for us.
We started on the registrar side of the problem. Study materials were fragmented across WhatsApp groups, private tutors, and imported MCQ banks that didn’t map to the CMSA blueprint. The hours were there; the right reference material was not.
- 02 / The observation
Frontier models are medically capable — in the wrong geography.
As soon as general-purpose LLMs became useful on clinical tasks, it was clear they defaulted to US and EU practice. Epinephrine instead of adrenaline. Trimethoprim instead of nitrofurantoin. Confident paragraphs where SA guidance was silent.
- 03 / The build
We built the model we wanted to use ourselves.
So we trained one — against SA Standard Treatment Guidelines, the tiered Essential Medicines List, IMCI, VTP, BANC+, the national TB algorithm. We built a benchmark to prove it moved. And we wrapped it in tools clinicians and trainees could actually reach.
- 04 / Today
Three surfaces, one model, one context.
Today Lorraine powers study with Learn, clinical reasoning support with Chat, and SA-aware integrations through the Platform API — all sitting on Lor-1, our clinical model for South Africa. Launched March 2026 and now in daily use by clinicians, registrars, and nurses across the country.
What we take a position on
Stances we won’t hedge on.
Four pairs that matter for how we build clinical AI — and why we’re on the side we’re on.
Our stance
At the weights, not the prompt.
What we reject
Wrap a foreign model in a prompt
System prompts and retrieval wrappers paper over drift. The moment context thins, the base reaches for the wrong default again.
What we do
Train the SA frame into the weights
SA drug names, tiered EML, SA STG, PHC → district → regional → tertiary become the model’s native reference — not a layer applied after the fact.