Inspiration

Healthcare planning is often forced to rely on incomplete, noisy, and inconsistent signals. We built Aarogya Atlas to help planners answer two questions together: where are the highest-risk care gaps, and how confident should we be in those gaps.

What it does

Aarogya Atlas is a transparent planning app that helps users identify and validate service gaps across geographies, review supporting evidence, and persist decisions for follow-up action.

Key features

Risk + confidence together: Ranks regions by care-gap urgency while showing confidence diagnostics, so users can distinguish real gaps from data-poor regions.
Evidence-backed outputs: Every major score and recommendation is traceable to underlying facility-level source text and clickable citations.
Uncertainty communication: Explicit flags and diagnostics explain where evidence is weak, partial, or suspicious instead of presenting low-quality data as fact.
Human-in-the-loop decisions: Users can override trust assessments with required notes, add facilities to shortlists, and save decision scenarios.
Persistent workflow state: Shortlists, overrides, and scenarios are saved and reloaded to support real planning workflows.
Geography-first exploration: State-led workflow with additional PIN/district views for deeper context.
Optional AI assistant: In-app Genie panel enables natural-language exploration while keeping deterministic scoring logic as the decision backbone.
Operational resilience: Dual persistence strategy (Lakebase primary with Delta fallback) plus health checks for warehouse/lakebase readiness.

How we built it

We deployed Aarogya Atlas as a Databricks App with a React frontend and AppKit backend.

Read/analytics path: Unity Catalog gold tables queried through Databricks SQL Warehouse
Write/transaction path: Lakebase Postgres for scenarios, overrides, and shortlist data
Reliability path: Delta fallback for persistence continuity during service interruptions
Trust model: Deterministic rule-based scoring and evidence signals for explainability

Challenges we ran into

State drill-down initially failed due to join-key formatting mismatches.
Citation fidelity required restructuring access patterns to preserve underlying text provenance.
Noisy text caused false positives and required tighter evidence guards.
Runtime service interruptions (warehouse/lakebase) required stronger health checks and recovery handling.
Fallback periods introduced split-storage reconciliation needs between Delta and Lakebase.

What we learned

Decision-support quality depends on transparency more than raw scoring power. The most important product outcome was making evidence quality, uncertainty, and provenance visible at every step, so users can make defensible decisions rather than just consume rankings.