Bharosa Care

BharosaCare: Transparent Healthcare Intelligence

Inspiration

Healthcare decisions shouldn't feel like a black box. When someone searches for care options—a specialist, a clinic, or treatment—they deserve to understand why they're seeing certain recommendations. We were inspired to build BharosaCare as a platform that harnesses the power of Databricks agents and Genie to make sense of messy, unharnessed healthcare data while giving users full transparency into how those AI-driven choices were made.

The name "Bharosa" means trust in Hindi and Urdu, and that's exactly what we set out to create: a trustworthy system where patients can explore care options near them with confidence, understanding the reasoning behind every suggestion.

How We Built It

BharosaCare is built entirely on Databricks:

Databricks Apps V2 for the user-facing interface and experience Genie and AI agents to parse and analyze fragmented healthcare datasets—provider ratings, insurance networks, location data, specialties, and availability Unity Catalog to organize and govern our data sources Lakebase (Unity Catalog Tables + Delta) for storing structured and semi-structured healthcare information We ingested real-world messy data—CSVs with inconsistent schemas, PDFs with provider information, and public datasets—and let our agents do the heavy lifting. The key innovation was building an explainability layer: every time Genie surfaces a recommendation, we capture the reasoning path (which data sources, which filters, which scores) and surface it directly in the UI.

What We Learned

Lakebase fundamentals: As beginners to Lakebase, we learned how to structure tables, leverage Delta Lake's ACID properties, and optimize for low-latency reads required by our app Serverless compute: We leaned on Databricks Serverless for SQL queries and Python jobs, which simplified our infrastructure Agent transparency: Building the explainability layer taught us how to instrument agent workflows—capturing metadata, lineage, and decision traces Rapid prototyping: Databricks Apps V2 let us iterate quickly from notebook experiments to a production-grade user experience

Challenges We Faced

Idle Lakebase compute: One of our biggest hurdles was managing compute resources effectively. Coming from a background of building Databricks apps but new to Lakebase-specific patterns, we struggled early on with compute clusters sitting idle when we expected them to be actively serving queries. We learned to:

Properly configure serverless SQL warehouses for our query patterns Use autoscaling and serverless compute to avoid idle time charges Optimize our Delta tables with Z-ordering and statistics to reduce compute overhead Data quality: Healthcare data is notoriously inconsistent—misspelled provider names, missing coordinates, outdated insurance info. We spent significant time building data quality rules and leveraging agents to auto-clean and standardize records.

Explainability UX: Translating agent decision paths into human-readable explanations was harder than expected. We iterated on the UI multiple times to find the right balance between detail and simplicity.

Built With

• databricks
• lakebase
• python
• streamlit