All projects
Felt cover for TriAIge

TriAIge

  • 1st place

A Turkish AI assistant that gives a safe pre-triage to people describing symptoms in their own words and routes them to the right specialty. It doesn't diagnose. An 8-layer pipeline catches emergencies right away. Comes with a patient app, a doctor portal and an embeddable clinic widget. First place at the AImpact Hackathon.

private codedots-hub

Award

First place at the AImpact Hackathon, with Triyaj.

My role: Three-person team: the backend and the AI system that keeps it consistent.

Problem

Many people with symptoms don't know where to go first. Some panic and go to the ER when they don't need to, others wait despite warning signs.

Approach

The person writes the complaint in their own words, and the app asks a few short questions and points to the right specialty. It doesn't diagnose. Emergency rules run before anything else, and no AI output can downgrade an emergency decision.

Result

A platform with a patient app, a doctor portal, an embeddable clinic widget and developer docs. All 165 cases in the under-triage regression suite pass. The project won first place at the AImpact Hackathon and is now in alpha.

Behind the scenes

Where the idea came from

In a disaster it is hard to reach care and to know what should come first. The idea came from a recent problem a teammate had with their hand. We wanted a system that looks at the symptoms and sends the person to the right place quickly.

The team

There were three of us, split into frontend, backend and data science. I had the backend and the AI system. We kept building after the hackathon ended.

The hardest decision

Not letting the AI diagnose. Fixed safety rules look at the critical symptoms first, and the language model only works in a supporting layer.

What we rebuilt

At first we depended on a single AI provider. We changed the architecture so that when a provider fails, the system moves to another provider or to the rule-based layer.

What the jury said

The jury said a system like this is needed in Turkey and around the world. Many jurors and other contestants offered to partner with us.

Where it stands

It is in alpha. We have not verified real user or clinic numbers yet. We have extensive automated tests, but they don't replace validation with real users.

If I started again

I would start clinical validation earlier and get feedback from health professionals sooner. I would build a small but reliable triage system first and add the advanced AI features later.

Highlights

  • 1st place, AImpact Hackathon
  • 8-layer pipeline
  • 165/165 under-triage cases
  • 6.5k backend tests

Stack

  • FastAPI
  • Python
  • Expo
  • React Native
  • Next.js
  • Supabase
  • PostgreSQL
  • multilingual-e5
  • Wiro
  • OpenAI
  • Anthropic
  • Railway
  • Sentry
  • Symptom entry screen
    The person describes the complaint in their own words.
  • Pain location question
    A narrowing question: where is the pain strongest?
  • Emergency screen
    On an emergency the flow stops and shows the 112 steps.