# VitaHub
co-founder, cto · 2024 – now
Software for health insurers' call centers. An operator types what the caller describes, answers a few follow-up questions, and gets the right specialty and a named doctor from the insurer's network, with the reason. The rule we built it around: software routes, doctors decide. I'm the CTO; the rest of the founding team are physicians.
It started with my mom. She had a problem with her finger and went from a rheumatologist to an orthopedist before she reached the right doctor. That is the whole problem: getting a patient to the right doctor the first time. I didn't know anyone in medicine, so I cold-emailed Dr. Jasmina Knežević, who founded and ran BelMedic hospital, and kept writing until she said yes.
## what i built
- The routing API: fixed rules for newborns, LLM follow-up questions, vector search over a hierarchy of 188 specialties, LLM ranking, then scoring of doctors in the client's network.
- The tool our physicians use to curate the doctor database (550+ doctors) and to write test cases.
- VitaEval, which replays labelled cases against the API and reports how often the right specialty comes back.
- The AWS setup: three accounts, single sign-on, deploys from GitHub with no stored keys, a private network behind a firewall.
- A co-pilot prototype that floats over the operator's screen as a transparent window and stays silent unless it can cite a source.
## what was hard
Not drifting into diagnosis. Under EU rules, software that diagnoses counts as a medical device. An audit I ran on our own API found 135 issues, 5 of them critical: it returned diagnosis codes, and fell back to a default code when the LLM failed. I rebuilt the API around a strict no-diagnosis boundary, with a test that fails if a diagnostic word reaches the output. The rewrite touched 263 files (+9.9k / −43.8k lines).
## what happened
- On our internal 40-case regression set: right specialty among the suggestions 38/40 → 40/40; cases with no doctor 4 → 0; median answer time 15.3 s → 8.8 s.
- Pitched at Goldman Sachs in October 2025, as part of Syracuse University's Entrepreneurship Intensive.
- A short trial at a private clinic; a white-label licence proposed to a major Serbian health insurer (no contract yet).
## links
# this site is a shell: open it here