facility-capability-trust

Inspiration India has thousands of healthcare facilities, but patients and care planners have no reliable way to understand which ones they can actually trust for a specific service. Confidence in a facility is often buried in data — clinical registries, government filings, inspection reports — that no patient can read or interpret. We wanted to change that: give patients and care planners a simple, honest view of which facilities have strong evidence behind them, and let that view improve over time through human review and AI-driven policy refinement.

What it does Facility Care Insights is a patient-facing healthcare facility trust platform. It lets users:

Browse 10,000+ India healthcare facilities filtered by service area, region, state, and city See each facility's trust score — a credibility rating derived from evidence signals like accreditation, inspection outcomes, and clinical registration data Read the supporting references behind each score so decisions are grounded in facts, not black boxes Add reviewer notes to adjust confidence when human judgment is needed View social proof: how many users increased or decreased their confidence in facilities this month Behind the scenes, a weekly agentic AI job (powered by Databricks Model Serving with databricks-gpt-oss-120b) reads reviewer notes, scoring drift, and error patterns from Lakebase Postgres, then generates policy delta proposals — candidate updates to the scoring rules — with rationale, expected impact, and risk notes. No policy changes production automatically; every proposal goes through human governance review.

How we built it Frontend: React + TypeScript + Tailwind CSS, served via Databricks AppKit Backend: Express.js server (TypeScript), deployed as a Databricks App Database: Databricks Lakebase (managed Postgres) — stores facility assessments, reviewer notes, scoring decisions audit log, policy versions, and AI-generated proposals AI / Agentic: Databricks Model Serving endpoint (databricks-gpt-oss-120b) called weekly via Databricks SDK from a serverless Lakeflow Job — reads context from Lakebase, generates structured JSON policy proposals Data: 10,000 curated India health facilities seeded from static catalog data with deterministic quality gating Infrastructure: Databricks Asset Bundles (databricks.yml) for job + app config; deployed via Databricks CLI Challenges we ran into Corporate network blocks: Zscaler intercepted workspace-files upload API calls during bundle deploy, requiring us to split the deploy process into sync + direct app deploy Serverless job dependencies: Databricks serverless jobs reject task-level libraries — dependencies must be declared in the environments block under spec.dependencies Serving SDK message format: The model serving SDK expected message objects with .as_dict(), not plain dicts — required a ChatMessagePayload wrapper class Browser caching of hashed assets: Vite's content-hashed bundles combined with Databricks Apps' snapshot model meant stale JS persisted until old hash files were explicitly deleted from the workspace Accomplishments that we're proud of A fully working end-to-end agentic loop: reviewer notes → Lakebase → weekly LLM job → structured policy proposal → governance review Patient-centered UX: non-technical language, social proof metrics, and explainable confidence labels that a patient (not just a data analyst) can act on 10,000 facility dataset with quality-gated seeding and deterministic deduplication, loaded into Lakebase at app startup A clean scoring governance framework with audit logs, drift detection views, and policy versioning — all in Postgres on Lakebase Successfully deployed a React + Express + Postgres + Agentic Job stack entirely on Databricks What we learned Databricks Lakebase is a strong fit for OLTP workloads in apps — token-based auth integrates cleanly with the SDK Serverless job environments require explicit dependency management in bundle config, not task-level libraries The workspace import-dir CLI encodes files as notebooks (base64) — for binary/JS assets, use databricks sync which uses the workspace-files API correctly Agentic proposals are most useful when the output schema is strict and the safety rules are explicit — LLMs without a strong schema tend to produce unusable policy diffs What's next for facility-capability-trust Live data pipeline: Replace static seed with a live Unity Catalog pipeline ingesting NHA, NABH, and state health department registries Policy proposal review UI: Let governance reviewers approve/reject AI proposals directly in the app Confidence trend charts: Show a facility's trust score over time as evidence and reviewer signals accumulate Patient feedback loop: Let patients who visited a facility rate their experience and feed that signal back into scoring Multi-country expansion: Extend the scoring framework to Southeast Asia and Sub-Saharan Africa facility registries

Built With

• app
• category-technology-language-typescript
• databricks
• databricks-appkit-database-databricks-lakebase-(managed-postgres-/-psycopg3)-ai-/-llm-databricks-model-serving-(databricks-gpt-oss-120b)-agentic-automation-databricks-lakeflow-jobs-(serverless)
• databricks-asset-bundles-(dabs)-cloud-aws-(via-databricks)-auth-databricks-oauth
• databricks-sdk-infrastructure-databricks-apps
• python-frontend-react-18
• shadcn/ui-backend-express.js
• tailwind-css
• token
• vite