Nemotron ranks; it writes none of the words a caregiver reads.
Inspiration
Heart failure, type 2 diabetes, and chronic kidney disease often occur together, and their guidelines share medicines and lab tests. Caregivers get a few minutes with the care team and leave with questions unasked. Navigator helps them arrive with the right questions, each traceable to a published guideline.
What it does
Three short steps: conditions, medicines, then test results and the main concern. The caregiver gets up to three questions, each in a guideline card's own wording with its DOI. They can reorder, remove, or add questions, then print a brief with "Before you go" prompts and an "After the visit" notes page. No login; nothing is saved. If the AI service is down, a labeled rule-based version still works.
How we built it
A starter set of 12 evidence cards from 5 published guidelines. Rules select the eligible cards. NVIDIA Nemotron (nvidia/nemotron-3-super-120b-a12b) on Nebius Token Factory, with JSON mode and thinking off, ranks them and flags matches to the concern. Test results never reach the model. Stack: React, TanStack Start, Tailwind, Bun, Vitest, on Lovable. Claude Code and Lovable helped write code; I made the product, content, and safety decisions.
Challenges we ran into
Prompting alone could not keep the model inside the evidence. Our audit found a crafted concern could make a drafted sentence reverse a card using only the card's words, which a word filter cannot catch. So the model now only ranks.
Accomplishments that we're proud of
Every word shown comes from a guideline card, with its source. Nemotron put the expected card in the top three for 7 of 7 test concerns; the rule-based fallback, 3 of 7.
What we learned
Word filters cannot see negation. The fix was removing the model's authorship, not filtering harder.
Evaluation
30 synthetic cases on the live site: Nemotron answered 28 of 30 (the other 2 had no eligible cards, by design). Median response 0.9 s, 90th percentile 1.1 s. 0 violations of 6 structural rules. 155 unit tests; accessibility scan clean. Small sample, not clinical validation.
What's next for Navigator by ConditionMap
Clinician review of the cards, a "what matters most" question with ValuesInCare, Spanish, and a small caregiver pilot.
Built With
- lovable
- nebius-token-factory
- nvidia-nemotron
- react
- tailwindcss
- tanstack-start
- typescript
Log in or sign up for Devpost to join the conversation.