Proven where it's hardest: what Türkiye taught our clinical AI
Building a medical AI is rarely about the model. It is about the world the model has to work in — a language, a set of codes, a rulebook. Here is why we chose the hardest version of that world, and what it taught us.
When people imagine building a medical AI, they usually picture the model first: the architecture, the parameter count, the benchmark scores. In practice, the model is rarely the hard part. The hard part is the world the model has to work in — a specific clinical language, a particular set of codes, and a thick rulebook that decides, clause by clause, whether a treatment is even payable. We decided to meet that world at its most demanding, and the lessons turned out to be far more useful than any single accuracy number.
Türkiye is, by most measures, a difficult place to start. Clinicians write in Turkish, but their notes are threaded with Latin terminology and local shorthand; the same complaint can appear as everyday speech in one file and as formal medical language in the next. Sitting on top of all that is the SGK and its SUT framework — the rulebook that governs, line by line, what social insurance will reimburse. A human operator needs years to internalize it. For a general-purpose model trained mostly on English web text, it may as well be a foreign country. That difficulty is exactly why we picked it: it is far easier to simplify a system that has already mastered complexity than to add complexity to one that never had to.
The approach has three layers, and it helps to understand what each one actually does. The first, continued pre-training, is less about teaching the model facts and more about raising it somewhere new. Instead of fine-tuning it against a checklist of instructions, we let it keep learning on Turkish medical terminology, clinical guidelines and real-world records until that language became, in effect, its mother tongue. The second layer adds specialty knowledge through small, swappable adapters — one for cardiology, one for coding, one for reimbursement — so a single model can act like many specialists without the cost of training many models. The third connects everything to current literature through retrieval, so an answer stays anchored to a source you can check rather than to the model's memory.
Did it work? In our internal evaluation, run across fourteen specialties on roughly 9,300 synthetic, anonymized cases, the model reached 87% top-1 accuracy on ICD-10 coding and 91% on SUT reimbursement decisions, and it cut the rate at which it invents or misattributes a regulatory clause to 0.3%. That last figure matters more than it first appears: in a reimbursement setting, a confidently wrong citation is not a rounding error — it is a rejected claim or an audit finding. And the trend was as telling as the totals. On a frozen test set, ICD-10 accuracy rose from 74% in v3.5 to 87% in v3.7 — steady, measurable progress rather than a lucky single run.
This is the part that turned a local project into a global thesis. The things that make Türkiye hard — a non-English clinical language, an intricate payer system, country-specific codes — are not Turkish problems. Every health system has its own version of each. And the method we used to absorb them is not tied to any one country: point the same three layers at a new market's language, rules and records, and the platform re-localizes. The difficulty was never the destination; it was proof that the vehicle travels.
It is worth being precise about what these results are and are not. They come from internal evaluation on synthetic data, with no real patient records, and the systems we compared against are leading general-purpose models, kept anonymous and run under the same conditions. That makes them a strong signal of capability on realistic tasks — not a substitute for prospective, real-world validation, which is the work we are doing now. A benchmark earns a pilot; a pilot earns a deployment. We would rather say that plainly than overstate it.
If there is one takeaway, it is this: in clinical AI, local competence is not a feature you bolt on at the end. It is the whole job. Starting where that job is hardest did not just produce better numbers — it gave us a method we can carry into the next market, and the one after that.