Indie Referral Pilot

Inspiration

Rural healthcare referral is often not a pure search problem. A patient or frontline worker may know symptoms and a village or PIN code, but not which facility can safely handle the need. The provided facility dataset is useful, but noisy: claims are uneven, fields are incomplete, and not every geography has trustworthy coverage.

We built Indie Referral Pilot to help a non-technical user ask: “Where should this patient go next?” The app is designed to be helpful while staying honest about uncertainty.

What it does

Indie Referral Pilot is a multilingual referral copilot for rural India. A user can describe symptoms and location in natural language, including Telugu, and the app resolves PIN code context, identifies care needs, searches curated Databricks gold tables, and returns referral guidance.

The app supports PIN code grounding, care-need search, evidence-attached facility recommendations, safe fallback guidance when verified local facility data is missing, emergency escalation, and NMC/APMC doctor registration context.

Doctor context is clearly labeled as registration metadata, not proof of current practice location or facility affiliation.

How we built it

We built the app on Databricks. Databricks Apps hosts the React and AppKit-based user experience. Delta tables store ingested and curated healthcare data. Lakeflow Jobs build the data model. Lakebase synced tables serve low-latency API queries. Databricks AI Functions enrich facility evidence into structured capability signals. OpenAI powers the multilingual conversational layer.

The data pipeline starts with the provided hackathon facility dataset and the India PIN code directory. We normalize messy facility evidence into gold tables for referral search, care-need taxonomy, facility evidence, geography coverage, village routing, and doctor search.

We also ingested Andhra Pradesh Medical Council records from the NMC registry to add registered doctor context. For enriched records, we pulled qualification and specialization details, such as Surgical Gastroenterology, and exposed them in the app with an explicit safety note.

Challenges we ran into

The biggest challenge was balancing usefulness and safety. The dataset does not always contain verified nearby facilities for a rural PIN code. Instead of hallucinating hospital names, the app explicitly says when verified local candidates are missing and recommends the safest next action.

Another challenge was syncing analytical Delta tables into an interactive serving path. We used Lakebase synced tables so the app could query gold data quickly while keeping the curation pipeline auditable.

The NMC data was also difficult to ingest at scale because the public endpoint was slow and intermittent. We made the exporter resumable, supported state-level extraction, and added a separate detail enrichment path for qualifications and specializations.

Accomplishments that we're proud of

We built an end-to-end Databricks application that goes from messy healthcare evidence to a working multilingual referral experience.

We are especially proud that the app refuses to hallucinate unsupported facility claims. It communicates uncertainty clearly while still giving practical next steps such as PHC, CHC, Area Hospital, ASHA/ANM, 104, or 108 ambulance.

We also added NMC doctor registration context and enriched specialist records while clearly separating registration evidence from current practice-location claims.

What we learned

We learned that trustworthy healthcare apps need more than an LLM answer. They need grounded data models, clear uncertainty boundaries, low-latency serving, and product decisions that protect users when evidence is incomplete.

Databricks worked well for this pattern: Delta for durable curation, Lakeflow Jobs for reproducible pipelines, AI Functions for evidence extraction, Lakebase for serving, and Databricks Apps for the final workflow.

What's next for Indie Referral Pilot

Next, we would add more verified village-level PHC/CHC and primary care center data, persist user shortlists and review notes, add human review workflows for suspicious facility claims, and expand doctor enrichment beyond Andhra Pradesh.

We attempted to source village-level primary care center information, but the available public source was restricted to Indian IP addresses during the hackathon. With access to that dataset, we would improve rural coverage by mapping villages to their nearest sub-centres, PHCs, CHCs, and referral hospitals.

We would also consider semantic search over facility evidence using Mosaic AI Vector Search once the evidence corpus is larger, and improve care-need-aware doctor ranking for specialties such as maternity, surgery, and gastroenterology.

Sample prompts for judges

Rural fever triage in Telugu

నాకు తలనొప్పి, కడుపు నొప్పి, జ్వరం ఉన్నాయి. నేను భీమలాపురం లో ఉన్నాను, 534266 పిన్‌కోడ్. దగ్గరలో ఎక్కడికి వెళ్లాలి?

Pregnant patient escalation in Telugu

నా భార్య గర్భిణి. ఆమెకు కడుపు నొప్పి, జ్వరం, బలహీనత ఉన్నాయి. మేము 534266 పిన్‌కోడ్ దగ్గర ఉన్నాం. వెంటనే ఎక్కడికి వెళ్లాలి?

Gastro specialist doctor context

నాకు gastro specialist doctor కావాలి. నేను Andhra Pradesh లో ఉన్నాను. రిజిస్టర్ అయిన డాక్టర్లు చూపించండి.

English fallback prompt

I have fever and stomach pain near PIN code 534266. What is the safest next step?

Note: For rural PIN code 534266, the app may intentionally say that verified nearby facility candidates are unavailable. That is expected: the app avoids hallucinating hospital names and instead routes safely to PHC/CHC/Area Hospital, ASHA/ANM, 104, or 108.

Built With

• databricks
• databricks-ai-functions
• databricks-appkit
• databricks-apps
• databricks-asset-bundles
• delta-lake
• express.js
• india-pin-code-directory
• lakebase-postgres-synced-tables
• lakeflow-jobs
• nmc-doctor-registry-api
• node.js
• openai-api
• postgresql
• react
• typescript
• unity-catalog