Inspiration
In Türkiye, most clinical knowledge sharing still happens in fragmented WhatsApp and Telegram groups. Valuable case discussions disappear in chat history, are impossible to search, and often include sensitive patient information shared over insecure channels. At the same time, there is almost no high‑quality, ethically collected Turkish medical data to power local AI models. BYZA was born from this gap: I wanted to build a platform that doctors would actually use in daily practice, while silently building the foundation for a Turkish medical LLM.
What it does
BYZA is a secure, KVKK‑compliant collaboration platform for healthcare professionals in Türkiye. It combines a Reddit‑style case discussion forum, in‑app secure messaging, and an AI layer that automatically anonymizes PHI/PII in Turkish clinical text and images. Doctors and students can discuss real cases, track patients, calculate clinical scores like NEWS2, and search similar cases via vector search — while the system continuously collects anonymized, quality‑weighted data to train a future Turkish medical language model.
How we built it
On the product side, BYZA consists of three main modules: a structured case forum (with templates like SOAP, Emergency, Procedure, Discharge), a secure chat system (1:1 and group messaging with WebSockets), and a patient tracking module with vitals, reminders, and reports. On the AI side, we built services for PHI/PII anonymization using Presidio + spaCy with 5 custom Turkish recognizers, image anonymization with EasyOCR + Tesseract, a keyword‑based ICD‑10 differential diagnosis assistant, and a MedicalLabeler service for medical text tagging. We use multilingual‑e5‑large embeddings and Qdrant for semantic similar‑case search. All consented, anonymized data flows into an ML training pipeline designed for SFT, DPO and corpus‑level training on top of a base model like Mistral‑7B with QLoRA.
Challenges we ran into
The hardest parts were aligning strict KVKK requirements with real‑world clinician workflows and building robust Turkish PHI/PII detection. General‑purpose NER models are not enough for Turkish medical context, so we had to design and test custom recognizers and pipelines. Another challenge was product UX: doctors are already overloaded, so the forum, chat, and patient‑tracking features needed to feel faster and more practical than their existing WhatsApp + Excel workflows, otherwise they would never switch. Finally, designing a data pipeline that is both privacy‑preserving and ML‑ready required careful schema and consent design.
Accomplishments that we're proud of
We shipped a production‑ready platform at https://byza.org with a full stack: case forum, secure messaging, patient tracking, gamification, admin tools, and AI services. We implemented a 10‑layer security architecture, diploma‑based user verification, and a granular consent system that lets users control how their data is used for AI training. On the AI side, we now have an end‑to‑end pipeline that can turn real Turkish clinical notes into anonymized training samples with quality scores, ready for SFT and preference‑based training of a Turkish medical LLM/SLM.
What we learned
We learned that “AI for healthcare” is not just about models; it is about integrating into existing clinician behavior with minimal friction and very strong privacy guarantees. We also saw how important it is to design data formats and consent flows from day one if you want to build a sustainable medical AI dataset. Finally, working with Turkish clinical language showed us how under‑served low‑resource medical languages are and how much opportunity there is to build local, trustworthy models.
What's next for Byza.ORG
Next, we plan to finalize Firebase push notifications and release the mobile apps to the App Store and Play Store to make BYZA accessible at the bedside. We will run a closed beta with medical students, residents, and specialists, then start hospital/faculty pilots to validate real‑world usage. In parallel, we will launch the first training cycles of a Turkish medical LLM/SLM on top of anonymized BYZA data, and expose parts of the anonymization and decision‑support stack as APIs for other healthcare institutions.
Log in or sign up for Devpost to join the conversation.