Case Study · Healthcare
Pusula Yazılım: AI in hospital software, five projects in six months
In the first half of 2025 we worked with the team at Pusula, a company that builds software for hospitals, and mentored five AI projects. A NoShow model that predicts whether a patient will attend their appointment, and language models that answer requests hospitals send through Jira and classify and route comments, went live. Reading lab test kits with image processing and turning patient–doctor consultations into the record format remained R&D prototypes.
- 6 months
- Duration
- 5
- Projects
- NoShow · Jira
- Live
- 2 prototypes
- R&D
January–June 2025
Appointments · lab · consultation · Jira
Appointment forecast and Jira requests
Lab analysis · patient–doctor consultation
Context
Pusula Yazılım builds software for hospitals: hospital ERP systems, an appointment management system and an app called Kendine İyi Bak. These products produce appointment, lab, consultation and support records every day. Pusula wanted to turn those records into AI features that run inside its products.
Between January and June 2025 Burak Suyunu worked alongside Pusula’s team. He mentored all five projects, from choosing the method to evaluating the result. We split the projects into two groups: those whose data was ready and could enter the product quickly, and those that first needed to be tried as a prototype.
What went live
NoShow: appointment forecasting. A model that predicts whether a patient will attend their appointment. A patient who does not show up means an empty hour for the doctor and the equipment. The forecast is used to make better use of hospital resources and to improve appointment planning.
Jira: answering requests and categorising comments. Hospitals send their requests to Pusula through Jira. Two models were built. The first answers incoming requests automatically and routes them to the right unit. The second sorts the comments in Jira into defined categories and passes them to the right person. A request now reaches the right place without anyone having to decide whose desk it lands on.
R&D prototypes
Lab test analysis. Reading the test kits used to detect urinary infections with image processing. The aim was to read the kit’s result quickly and consistently and shorten the diagnosis process.
Patient–doctor consultation. Analysing the conversation between patient and doctor and turning it into the record format set by the hospital. When the record is not written by hand, time is saved and records accumulate in the same structure.
Both were built as prototypes; neither was taken into the product.
Language models by two routes
In the language projects we used two routes side by side: large language models through APIs, and local models we fine-tuned. The API makes quick experiments possible. A local model keeps the data inside the institution, which with patient data is a requirement, not a preference. All development was done in Python.
Outcome
At the end of six months two pieces of work were live in Pusula’s products: the appointment no-show forecast and the answering and routing of Jira requests. On the image processing and conversation analysis side, the team was left with working prototypes that can later be taken into the product.