Skip to content

petexp/HackMITPub

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Referral Copilot

Evidence-attached, trust-scored facility shortlists for Indian healthcare. Built for Challenge 04 "Data Legend" — Databricks × Hack-Nation, 6th Global AI Hackathon.

In India, a postal code can determine a lifespan. Families travel hours to a hospital only to find the ICU was a claim, not a capability. Referral Copilot helps a non-technical NGO planner answer one question — "where should this patient actually go?" — with an answer they can defend.

Type a location and a care need ("dialysis near Jaipur") and you get a ranked shortlist of facilities. Every candidate shows:

  • Distance to the patient's location,
  • the verbatim sentence from the facility's own record that supports the match (never paraphrased — citations are enforced as exact substrings),
  • the data gaps we couldn't fill,
  • a calibrated trust score with an uncertainty interval, and
  • automatic consistency flags (e.g. "ICU" claimed with no critical-care evidence).

You can override any assessment with a note and save the shortlist — and it survives the session.

Why it's different

  • Evidence, enforced in code. Every citation must be a verbatim substring of the source text, checked on every path — so a model can't hallucinate a capability into the results.
  • Confidence without a ground truth. A deterministic, fully explainable trust score separates strong evidence from weak claims and reports how sure it is.
  • It double-checks its own work. A rule-based validator penalizes internally inconsistent claims and shows why.
  • A data desert is not a medical desert. Missing data is shown as a gap, never rendered as "no facilities."

Tech at a glance

A seven-stage pipeline — parse → geocode → retrieve → extract → trust → validate → rank — over a medallion lakehouse built from ~10,000 facility records. It runs fully offline on the bundled sample (BM25 retrieval, keyword extraction, SQLite), and transparently upgrades to Databricks services when configured: Unity Catalog, Mosaic AI Vector Search, Foundation Models, Lakebase, a Serverless SQL warehouse, and MLflow tracing — shipped as a Databricks App.

Path Purpose
app/ Streamlit UI — search, ranked cards, map, saved shortlists
copilot/ UI-free business logic (the pipeline stages), fully unit-tested
notebooks/ Medallion data pipeline (runs locally and as Databricks notebooks)
sql/ Databricks SQL artifacts (Lakebase schema, etc.)
data/ Small seed data + reference lookups (large datasets are gitignored)
tests/ pytest suite

Quickstart

Requires Python ≥ 3.10.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

ruff check . && pytest          # lint + tests
streamlit run app/app.py        # run the app

The app opens on the Referral tab. Click a preset (Dialysis near Jaipur / Emergency surgery near Patna) or enter your own location + care need, and hit Search. No API keys or network connection are required to run the demo.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages