Local AI for
hospital meetings
Challenge. Hospital boards at Medpark end with someone replaying the audio to write the Minutes of Meeting: decisions, actions, owners, deadlines. No clinic can afford to hand that data to a cloud service, so it all had to stay inside the hospital network.
Research. No local AI yet understands Moldovan doctors mixing Romanian, Russian and English with mixed accents and medical terms, so I listened to a real Medpark meeting and mapped 12 patterns for our AI to learn.
Role map. One recording, three wins: the manager sends the MoM faster, doctors get their actions sooner, the hospital saves time and keeps data inside. Because of the time limit and no access to doctors, I mapped value on one sticky board instead of user interviews.
Decisions. I turned the role map stickers into 8 UI rules, so the meeting manager no longer spends 2 hours writing the MoM by hand, but just checks the actions in a few minutes, confirms or edits, and sends.
Process. AI cut the time from idea to usability testing. Prototype 1, built straight in code, let developers train the AI engine meanwhile I collected feedback from Medpark and improved the design over several rounds of iterations, up to the handoff version.
User flow. The only loops are safety checks: edit a wrong action, name an unknown voice. Nothing is sent until every action is confirmed and every participant has an email.
Design system. Generated with AI from a few simple rules: one action colour, 4 core text sizes, 4-pt grid.
Interface. One screen, one goal, one blue button. AI writes the MoM, the manager only checks and confirms.
Screen recording · main flow · one take, ~30–40 s
Start → Recording → Processing → Check MoM → edit an action item → Preview MoM → Sent
AI draft vs Design. AI gave a working screen fast. I decided what to remove, what to keep, and made the review screen earn trust.
Prototype 1
Prototype 2
Prototype 1
Prototype 2
After the demo. Our team got the second place. At a deep tech hackathon, the score leaned on security, linguistic accuracy and output quality, where the winning team's pipeline led. UX counted for only 10%.
