Inspiration
Millions of people manage an aging parent's care on top of their own lives. The records exist, but they're written for doctors and billing offices: ICD-10 codes, clinical shorthand, scanned lab reports, and bills with no explanation. Caregivers miss refills, pay bills they should question, and walk into appointments unprepared.
What it does
Care Compass reads one patient's whole record and tells the caregiver what matters. My persona is Maria, caring for her 78-year-old dad Joseph, who has 8 ongoing conditions and several doctors.
- Needs your attention: medications that have run out and what to do about each, pharmacist notes that need action, bills well above the usual charge, and claims still with insurance.
- Plain-language explanations of his conditions, each visit's doctor notes, and scanned documents like lab reports.
- Bill review: every bill is compared with the typical charge for the same visit type and procedure code across 164,695 claims. Unusually high bills get a one-click, ready-to-edit letter to the billing office.
- Appointment prep: questions to ask, tied to his actual record.
- Clinical trials related to his conditions, with what each would involve.
- Ask about Dad: a chat answered only from his record.
It adapts to the user. The caregiver sets their name, their relationship to the patient, their background (new to all this, or some medical training) and their language (English, Spanish, Arabic, Polish or Chinese). Every explanation changes to match; letters to billing offices stay in English. It explains, but never diagnoses or suggests changing treatment.
How I built it
Everything runs inside Snowflake:
- SQL data layer: one script joins 12 healthcare tables into app-ready tables in my own schema, using window functions to pick the 200 richest patient records and percentile functions to build the typical-charge benchmark.
- Semi-structured data: vital signs and trial protocols stored as JSON (VARIANT) feed blood pressure and weight trends and the trial explanations.
- Streamlit in Snowflake (warehouse runtime) serves the app, with its code on an internal stage, so the data never leaves the platform.
- Cortex AI (COMPLETE with Claude Sonnet 4.6) reads clinical notes, OCR'd documents, pharmacist notes and trial protocols, then explains, drafts letters and answers questions, with guardrails in every prompt.
Challenges I ran into
- The source data is admin-owned and read-only, so I used secondary roles to read it while keeping every app object owned by my own role.
- New Streamlit apps defaulted to a container runtime that couldn't run my code, so I created the app with SQL on the warehouse runtime.
- The warehouse runtime ships Streamlit 1.22, so I wrote compatibility fallbacks for newer features.
- Synthetic data assigns genders at random, so pronouns follow the relationship the caregiver picks, not the record.
Accomplishments that I'm proud of
- A real end user, not a data team, can get value from it in minutes.
- The bill benchmark is a genuine analysis: 162 of the 1,882 bills for the 200 patients in the app stand out, some at nearly double the usual charge.
- The same record adapts to different readers: a daughter, a Spanish-speaking parent, or a home health aide.
What I learned
AI is most useful here when it reads the unstructured parts (doctor's notes, scanned reports, pharmacy notes) that a caregiver can't, and when every answer is grounded in the record and limited to explaining.
How baked is it: 3, working prototype
It runs end to end on the event's synthetic data with live Cortex calls, for any of 200 patients, with no mock-ups or canned answers. Rough edges: no login, so anyone with the app can pick any patient; "today" is the latest date in the records; current medications are inferred from prescriptions; trial matching is by condition category, not eligibility rules; and error handling for slow AI calls is basic.
What's next for Care Compass
Caregiver logins tied to one family member, real eligibility checks for trials, refill reminders by text or email, and connecting to real patient portals with the patient's consent.
Log in or sign up for Devpost to join the conversation.